Blind image quality assessment (BIQA), which aims to accurately predict the
image quality without any pristine reference information, has been highly
concerned in the past decades. Especially, with the help of deep neural
networks, great progress has been achieved so far. However, it remains less
investigated on BIQA for night-time images (NTIs) which usually suffer from
complicated authentic distortions such as reduced visibility, low contrast,
additive noises, and color distortions. These diverse authentic degradations
particularly challenges the design of effective deep neural network for blind
NTI quality evaluation (NTIQE). In this paper, we propose a novel deep
decomposition and bilinear pooling network (DDB-Net) to better address this
issue. The DDB-Net contains three modules, i.e., an image decomposition module,
a feature encoding module, and a bilinear pooling module. The image
decomposition module is inspired by the Retinex theory and involves decoupling
the input NTI into an illumination layer component responsible for illumination
information and a reflectance layer component responsible for content
information. Then, the feature encoding module involves learning multi-scale
feature representations of degradations that are rooted in the two decoupled
components separately. Finally, by modeling illumination-related and
content-related degradations as two-factor variations, the two multi-scale
feature sets are bilinearly pooled and concatenated together to form a unified
representation for quality prediction. The superiority of the proposed DDB-Net
is well validated by extensive experiments on two publicly available night-time
image databases.
Abstract—Blind image quality assessment (BIQA), which aims to accurately predict the image quality without any pristine reference information, has been highly concerned in the past decades.
Especially, with the help of deep neural networks, great progress has been achieved so far.
特に、ディープニューラルネットワークの助けを借りて、これまで大きな進歩を遂げてきた。
0.71
However, it remains less investigated on BIQA for night-time images (NTIs) which usually suffer from complicated authentic distortions such as reduced visibility, low contrast, additive noises, and color distortions.
The image decomposition module is inspired by the Retinex theory and involves decoupling the input NTI into an illumination layer component responsible for illumination information and a reflectance layer component responsible for content information.
Then, the feature encoding module involves learning multi-scale feature representations of degradations that are rooted in the two decoupled components separately.
Finally, by modeling illumination-related and contentrelated degradations as two-factor variations, the two multiscale feature sets are bilinearly pooled and concatenated together to form a unified representation for quality prediction.
D UE to the poor lighting condition in night-time,
d 夜間の照明条件の悪さについて
0.56
the captured night-time images (NTIs) are usually perceived with poor visibility and low visual quality.
撮影された夜間画像(ntis)は通常、視認性が悪く視覚品質が低いと知覚される。
0.66
Given that highquality NTIs are crucial for consumer photography and practical applications such as automated driving systems, many NTI quality/visibility enhancement algorithms have been proposed.
However, the research efforts on designing objective quality metrics that can automatically quantify the visual quality of NTIs and compare the performance of different NTI enhancement algorithms remain limited, which hereby hinders the development of this field.
Generally, objective image quality assessment (IQA) methods can be roughly divided into three categories, i.e., full-reference (FR), no-reference (NR),
Q. Jiang and J. Xu are with the School of Information Science and Engineering, Ningbo University, Ningbo 315211, China (e-mail: jiangqiuping@nbu.edu .cn).
W. Zhou is with the Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, Canada (e-mail: wei.zhou@uwaterloo.c a).
W. Zhouはカナダのウォータールー大学電気・コンピュータ工学科(Eメール:wei.zhou@uwaterloo. ca)に所属している。
0.78
X. Min and G. Zhai are with the Institute of Image Communication and Information Processing, Shanghai Jiao Tong University, Shanghai 200240, China (minxiongkuo, zhaiguangtao@sjtu.ed u.cn).
Among them, FR and RR IQA methods require full and partial reference information, respectively.
FRとRR IQAは、それぞれ完全な参照情報と部分参照情報を必要とする。
0.75
However, for the NTIs we concerned, there is usually no available pristine image to provide any reference information.
しかし、我々が懸念しているNTIに対して、通常、参照情報を提供するプリスタンイメージは存在しない。
0.67
Therefore, NR-IQA is more valuable for NTIs in this regard.
したがって、NR-IQAはNTIにとってより貴重である。
0.71
Early studies on NR-IQA mainly focus on specific distortion types, i.e., assuming that a particular distortion type is known and then specific distortion-related features are extracted to predict image quality [2–6].
Obviously, such the specificity limits their applications like the real-world night-time scenario.
明らかにそのような特異性は、現実世界の夜間シナリオのようなアプリケーションを制限する。
0.63
Although the rapid advances in the IQA community during the last decade push to produce general-purpose blind IQA (BIQA) methods [7–24] that can simultaneously work with a number of distortion types, their efficacies are still limited to synthetic distortions.
This is evident by the fact that they usually validate their performance on legacy synthetic distortion benchmark databases where the distorted images are simulated from pristine corpus in laboratory.
As a result, the existing general-purpose BIQA methods still cannot work well with the authentically distorted images like the NTIs captured in the real-world night-time scenario.
Recently, inspired by the success of deep neural networks in many image processing and computer vision tasks, great progresses have also been achieved on deep learning-based BIQA.
However, it remains less investigated on deep learning-based BIQA for NTIs which usually suffer from complicated authentic distortions such as reduced visibility, low contrast, additive noises, invisible details, and color distortions.
The diverse authentic degradations in NTIs pose great challenges to the design of highly effective end-to-end deep network architectures for blind NTI quality evaluation (NTIQE).
To evaluate the visual quality of NTIs, Xiang et al [25] first established a dedicated large-scale natural NTI database (NNID), which contains 2, 240 NTIs with 448 different image contents captured by three different photographic equipments in real-world scenarios along with their corresponding subjective quality scores (obtained by conducting human subjective experiments).
NTIの視覚的品質を評価するため、Xiangらはまず2,240のNTIと448の異なる448の異なる画像コンテンツと、対応する主観的品質スコア(人間の主観的実験によって得られる)を含む、大規模なNTIデータベース(NNID)を構築した。 訳抜け防止モード: NTIの視覚的品質を評価する。 Xiang et al [ 25 ] は最初,大規模なNTIデータベース(NNID)を専用に構築した。 実世界の3つの異なる写真機器が捉えた448種類の画像を含む240のNTIと、それに対応する主観的品質スコア(人間の主観的実験によって得られる)を含む。
0.75
Then, a NR quality metric called BNBT is proposed by considering both brightness and texture features.
そこで, BNBTと呼ばれるNR品質指標を提案し, 明るさとテクスチャ特性の両面から検討した。
0.63
The experimental results on NNID database have demonstrated an acceptable performance of BNBT, i.e., the predicted quality scores by BNBT are consistent with ground truth subjective quality scores.
Despite its effectiveness, BNBT requires elaborately-designed handcrafted features, which enlightens us to adopt an end-to-end data-driven method by taking the advantage of deep learning.
It contains an image decomposition module, a feature encoding module, and a bilinear pooling module.
画像分解モジュール、特徴エンコーディングモジュール、バイリニアプーリングモジュールが含まれている。
0.66
The image decomposition module takes an NTI as input and decouples it into two layer components, i.e., illumination (L) and reflectance (R).
画像分解モジュールは、NTIを入力として取り、照明(L)と反射(R)の2つの層成分に分解する。
0.76
Then, the feature encoding module involves learning multi-scale feature representations of degradations that are rooted in the illumination and reflectance separately.
An important observation is that the commonly-encountered distortions in NTIs can have impacts on either illumination perception or content perception.
NTIにおける一般的な歪みは、照明知覚またはコンテンツ知覚に影響を及ぼす可能性がある。
0.74
For example, the color distortion and additive noise are only influencial in content perception while the reduced visibility and low contrast are only influencial in illumination perception.
Thus, it is intuitive to consider decomposing the input NTI into two independent components with each component accounting for illumination information and content information, respectively.
Assisted by such a tailored image decomposition process, the degradation features related to illumination perception and content perception can be better learned and then fused to facilitate blind NTIQE.
Inspired by the Retinex theory [26], the image decomposition module involves decoupling the input NTI into two layer components, i.e., one layer component (illumination) is responsible for illumination information, while the other one (reflectance) for content information.
retinex theory [26]にインスパイアされた画像分解モジュールは、入力ntiを2つのレイヤコンポーネントに分離すること、すなわち、一方のレイヤコンポーネント(照明)が照明情報を担当し、もう一方のレイヤ(反射)がコンテンツ情報を扱う。
0.77
Then, the feature encoding module involves learning multi-scale feature representations of degradations that are rooted in the two decoupled components separately.
and content-related degradations as two-factor variations, the two multi-scale feature sets are bilinearly pooled and concatenated together to form a unified representation for quality prediction.
Extensive experiments conducted on two publicly available night-time image databases well demonstrated the superiority of the proposed DDB-Net against state-of-the-art BIQA methods.
In summary, this paper presents the following contributions:
まとめると,本稿では以下の貢献について述べる。
0.64
1) We make the first attempt to perform Retinex decomposition to facilitate NTIQE by decoupling the input NTI into two independent layer components (i.e., illumination and reflectance) with each component accounting for illumination information and content information, respectively.
2) We introduce a self-reconstruction- based feature encoding module and design tailored loss functions to regularize the training process towards learning multi-scale illunimationrelated and content-related feature representations from the two decoupled components separately.
3) We model the illumination-related and content-related degradations as two-factor variations and perform bilinear pooling to fuse the two multi-scale feature sets into a unified representation for quality prediction of NTIs.
The rest of this paper is organized in the following manner.
この論文の残りは以下のように整理されている。
0.71
Section II introduces the related works.
第2節では関連作品を紹介する。
0.45
Sections III illustrates the proposed method with details.
第3節では、提案手法を詳述する。
0.55
Section IV presents the experimental results.
第4節は実験結果を示す。
0.65
Section V concludes the paper.
第5節はその論文を締めくくる。
0.39
In this section, we will review the existing related works, including traditional blind image quality assessment, deep
本節では,従来のブラインド画像品質評価,deepなど,既存の関連作品について概説する。
0.71
II. RELATED WORKS R𝑹ℒ𝑟𝑒𝑐𝑅=1−𝑆𝑆𝐼𝑀(𝑅,𝑅)+𝑅𝐵−𝑅𝐵22L𝑳ℒ𝑟𝑒𝑐𝐿=𝐿−𝐿22Bilinear PoolingConv_32_1×1_1Conv_32_1×1_1FCFCQConv_3×3Conv_3×3+ReLUConv_3×3+ReLU+MaxPooling_2×2Conv_3×3+ReLU+Upsampling_2SigmoidC opyConcatenate1232Im age Decomposition ModuleFeature Encoding ModuleFeature Encoding ModuleBilinear Pooling Module
learning-based blind image quality assessment, and blind image quality assessment in poor conditions.
学習に基づくブラインド画像品質評価、およびブラインド画像品質評価。
0.51
A. Traditional Blind Image Quality Assessment
a. 伝統的ブラインド画像品質評価
0.81
In the literature of traditional blind image quality assessment, natural scene statistics (NSS) and human visual system (HVS) are two main cues for designing objective BIQA models.
As for NSS-based frameworks, Moorthy et al [10] proposed the Distortion Identification-based Image Verity and INtegrity Evaluation (DIIVINE) index to evaluate perceptual image quality in a no-reference manner, which is composed of distortion identification and NSS-based quality regression.
nssベースのフレームワークに関して、moorthy et al [10]は歪み識別とnssに基づく品質回帰からなる知覚的画像品質を評価するために、歪み識別に基づく画像妥当性・完全性評価(diivine)インデックスを提案した。
0.74
Likewise, the CurveletQA [9] also performs within a two-stage framework containing distortion classification and quality assessment.
同様に、CurveletQA[9]も歪み分類と品質評価を含む2段階のフレームワークで実行されます。
0.80
Different from DIIVINE, the quality assessment of CurveletQA is based on NSS features in curvelet domain.
In addition, some opinion-unaware BIQA methods based on NSS, i.e. so-called “completely blind” models, such as Natural Image Quality Evaluator (NIQE) [18] and ILNIQE [17], have shown competitive performance with the help of massive natural images.
Beyond the NSS features, many other statistical factors have been considered by researchers.
NSSの特徴以外にも、多くの統計要因が研究者によって検討されている。
0.62
In the family of twostage framework, Liu et al [15] proposed the Spatial-Spectral Entropy-based Quality (SSEQ) index, where local spatial and spectral entropy features are used to predict perceptual image quality.
2段階の枠組みの族において、liu et al [15]は空間スペクトルエントロピーベース品質(sseq)指標を提案し、局所空間エントロピー特徴とスペクトルエントロピー特徴を用いて知覚画像品質を予測する。
0.74
The GM-LOG method [13] extracts the joint statistics of local contrast features to assess image quality, including gradient magnitude and Laplacian of Gaussian response.
Among these methods, Gu et al [12] proposed the Noreference Free Energy-based Robust Metric (NFERM) on the basis of free energy principle.
これらの方法のうち、Gu et al [12]は、自由エネルギー原理に基づいて、NFERM(Noreference Free Energy-based Robust Metric)を提案した。
0.74
In [19], Li et al used contrast masking to design the BIQA model based on structural degradation.
19]では、Liらはコントラストマスキングを用いて構造劣化に基づくBIQAモデルを設計した。
0.74
Besides, according to the similar concept regarding the HVS properties, they proposed the GWH-GLBP by computing the gradient-weighted histogram of local binary pattern [14].
However, the above-mentioned conventional BIQA methods generally need to design elaborate handcrafted features with the pre-defined NSS or HVS mechanisms.
Additionally, Bosse et al [35] presented the end-to-end WaDIQaM that can blindly learn perceptual image quality.
さらにbosse et al [35]は、知覚的な画像品質を盲目的に学習できるエンドツーエンドのwadiqamを発表した。 訳抜け防止モード: さらに、Bosse et al [35 ] が最後に提示した。 -to - end WaDIQaM 視覚的なイメージの質を 学べます
0.61
The deep bilinear convolutional neural network called DBCNN was proposed to bilinearly pool the feature representations to a single quality score [36].
This is mainly because these models usually neglect the specific characteristics of NTIs, e g reduced visibility, low contrast, additive noises, invisible details, and color distortions.
In such poor conditions, capturing images with high-quality is quite challenging and thus addressing the blind quality assessment issue is urgently needed.
In the quality evaluation of hazy images, Min et al [37] proposed the haze-removing features, structure-preserving features, and over-enhancement features to construct the objective quality assessment index.
hazyイメージの品質評価において、min et al [37]は、客観的品質評価指標を構築するために、haze除去機能、構造保存機能、過剰強調機能を提案した。 訳抜け防止モード: hazyイメージの品質評価において、min et al [37 ]はhaze - remove features を提案した。 構造 - 特徴の保存と拡張機能 客観的品質評価指標を構築する。
0.71
They also used synthetic hazy images to build an effective quality assessment model for image dehazing [38].
For image deraining, an efficient objective quality assessment model to predict the human perception towards derained images was developed, which belongs to a bi-directional gated fusion network [39].
They further extended it to a bi-directional feature embedding network to further advance the performance [40].
さらに、パフォーマンスをさらに進めるために、双方向の機能埋め込みネットワークに拡張した[40]。
0.77
As for underwater image quality evaluation, Yang et al [41] proposed to linearly combine chroma, saturation and contrast factors for quantifying the perceptual quality of underwater images.
Finally, we present the bilinear pooling module for fusing the two hierarchical feature sets.
最後に、2つの階層的特徴集合を融合させる双線型プールモジュールを提案する。
0.64
A. Image Decomposition Module
A.画像分解モジュール
0.78
4 According to the Retinex theory [26], a single image I can be considered as a composition of two independent layer components, i.e., reflectance R and illumination L, in the fashion of I = R ⊗ L, where ⊗ denotes element-wise product.
4 レチネックス理論 [26] によれば、単一の像 I は独立な2つの層成分、すなわち反射率 R と照明量 L の合成とみなすことができる。 訳抜け防止モード: 4 Retinex 理論 [26 ] によれば、単一の像 I を2つの独立した層成分の合成と考えることができる。 すなわち、反射率Rと照明量Lは、I = は元を表し、右積は右積である。
0.53
However, recovering two independent components from one single input is a typical ill-posed problem.
しかし、1つの入力から2つの独立したコンポーネントを復元することは、典型的な誤った問題である。
0.51
Here, we resort to deep neural network to achieve this goal.
ここでは、この目標を達成するためにディープニューラルネットワークを使用します。
0.61
In what follows, we first describe the detailed architecture of our image decomposition module and then present how to train it in advance.
That is, the image decomposition module is pretrained and kept fixed during the training of DDB-Net.
すなわち、画像分解モジュールは事前トレーニングされ、DDB-Netのトレーニング中に固定される。
0.74
The architecture of image decomposition module is shown in the upper left of Fig 1.
画像分解モジュールのアーキテクチャは、図1の上左に示される。
0.73
It contains two streams corresponding to the reflectance (R) and illumination (L), respectively.
それぞれ反射(R)と照明(L)に対応する2つのストリームを含む。
0.79
The reflectance stream adopts a typical 5-layer U-Net, followed by two convolutional (conv) layers and a Sigmoid layer in the end, while the illumination stream is composed of two conv+ReLU layers and a conv layer on concatenated feature maps from the reflectance branch, finally followed by a Sigmoid layer in the end.
Since no/few ground-truth reflectance and illumination maps for real images are available, designing a well-defined nonreference loss function is the key to the success for training a robust deep Retinex decomposition network.
This inspires us to take a pair of images (describing the same scene) as input and impose both reflectance and illumination constraints between the image pair to train the image decomposition module.
Specifically, during the training stage, the input to image decomposition module is an image pair of the same scene with different light/exposure configurations, as denoted by [Il, Ih].
具体的には、トレーニング段階において、画像分解モジュールへの入力は、[Il,Ih]で示されるように、異なる光/露光構成の同じシーンのイメージ対である。 訳抜け防止モード: 具体的には、トレーニング段階では、画像分解モジュールへの入力は、異なる光/露光構成の同じシーンのイメージ対である。 Il , Ih ] で表されるように.
0.78
Similarly, the decomposed reflectance and illumination components are denoted by [Rl, Rh] and [Ll, Lh], respectively.
同様に、分解された反射率と照明成分はそれぞれ [Rl,Rh] と [Ll,Lh] で表される。
0.84
The training of the proposed image decomposition module is guided by hybrid loss terms which are to be detailed subsequently.
提案した画像分解モジュールのトレーニングは、その後詳細となるハイブリッド損失項によって導かれる。
0.76
Inter-consistency loss: The inter-consistency loss includes reflectance consistency loss and illumination mutual consistency loss.
相互整合損失: 相互整合損失は、反射率整合損失と照明相互整合損失を含む。
0.66
First, the reflectance consistency loss LR con encourages the reflectance similarity, which is defined as follows:
まず、反射率一貫性損失lr conは反射率類似性を助長し、次のように定義する。
0.65
(1) where (cid:107) · (cid:107)1 means the (cid:96)1 norm.
1) (cid:107) · (cid:107)1 は (cid:96)1 ノルムを意味する。
0.70
Second, the illumination mutual consistency loss LL
第2に、照明相互整合損失LL
0.70
con = (cid:107)Rl − Rh(cid:107)1, LR (cid:18)
con = (cid:107)Rl − Rh(cid:107)1, LR (cid:18)
0.44
con is defined as follows: c2 ⊗ exp
con は次のように定義される。
0.62
− M 2 2c2 LL con = f (M ) =
-M22c2 LL con = f (M ) =
0.39
(cid:13)(cid:13)(cid :13)(cid:13) M
(cid:13)(cid:13)(cid :13)m
0.46
(cid:19)(cid:13)(cid :13)(cid:13)(cid:13) 1
(cid:19)(cid:13)(cid :13)(cid:13)1
0.45
, (2) Image reconstruction loss: The third consideration is that the decomposed two components should well reproduce
Fig. 2. Penalty curves with different values of c.
図2。 c の値が異なるペナルティ曲線。
0.61
M = |(cid:79)Ll| + |(cid:79)Lh| ,
M = |(cid:79)Ll| + |(cid:79)Lh| 。
0.79
(3) where (cid:79) means the first order derivative operator along both horizontal and vertical directions, c is a parameter controlling the shape of the above penalty curve.
To facilitate understanding, we draw the penalty curves with different values of c in Fig. 2.
理解を容易にするために、図2でcの値が異なるペナルティ曲線を描く。
0.71
As we can see, the penalty value first increases and then decreases to zero as M increases.
ご覧の通り、ペナルティ値が最初に増加し、Mが増加するにつれて0に減少する。
0.71
In our implementation, we set c = 0.1 empirically.
実装では、c = 0.1 を経験的に設定する。
0.62
By minimizing such an illumination mutual consistency loss, the mutual strong edges are encouraged to be well preserved and all weak edges are to be suppressed.
Individual smoothness loss: Besides the inter-consistency loss, we also consider isolate loss for each decomposed component separately by considering their own smoothness properties.
On the one hand, the illumination maps should be piece-wise smooth, thus we introduce a structure-aware smoothness loss LL S to constraint both Ll and Lh: (cid:79)Ll
一方、照明写像は断片的に滑らかでなければならないので、Ll と Lh: (cid:79)Ll を制約するために構造対応の滑らかさ損失 LL S を導入する。
0.74
(cid:79)Lh
(cid:79)lh
0.38
max{((cid:79)Rl)2, τ}
max{((cid:79)Rl)2, τ}
0.49
+ max{((cid:79)Rh)2, τ}
+ max{((cid:79)Rh)2, τ}
0.46
(cid:13)(cid:13)(cid :13)(cid:13)1
(cid:13)(cid:13)(cid :13)1
0.48
(cid:13)(cid:13)(cid :13)(cid:13)
(cid:13)(cid:13)(cid :13)
0.47
(cid:13)(cid:13)(cid :13)(cid:13)1
(cid:13)(cid:13)(cid :13)1
0.48
, (4) (cid:13)(cid:13)(cid :13)(cid:13)
, (4) (cid:13)(cid:13)(cid :13)
0.44
LL sm = to the reflectance.
LL sm = 反射率に
0.46
Therefore, where τ denotes a small positive constant which is empirically set to τ = 0.01 to avoid the denominator being zero.
そのため ここで τ は τ = 0.01 に実験的に設定された小さな正の定数を表す。
0.74
This loss measures the relative structure of the illumination with respect the illumination loss can be aware of image structure reflected by the reflectance.
Specifically, for a strong edge point in the reflectance map, the penalty on the illumination will be small; for a point in the flat region of the reflectance map, the penalty on the illumination turns to be large.
On the other hand, different from the illumination maps that should be piece-wise smooth, the reflectance maps are usually tend to be piece-wise continuous.
(7) Fig. 4. Architecture of our self-reconstruction- based encoder-decoder network for hierarchical feature learning.
(7) 図4。 階層的特徴学習のための自己再構成型エンコーダ・デコーダネットワークのアーキテクチャ
0.57
To facilitate understanding, as shown in Fig 3, we draw a simple diagram to better illustrate the training process of our image decomposition module where all the involved loss terms are specified.
It should be emphasized that our image decomposition module is pre-trained on a collection of images pairs of the same scenes with different light/exposure configurations.
After obtaining the reflectance and illumination components via the pre-trained image decomposition module, the next step is to build feature representations for each of these two components separately.
However, these two feature encoding networks are optimized with different loss terms, i.e., tailored loss terms are designed to regularize the feature encoding of reflectance and illumination components separately.
The encoder receives either the reflectance (R) or illumination (L) component as input and progressively forms a set of hierarchical feature representations C1, C2, C3, C4, and C5.
Since the reflectance and illumination components contain NTI degradation information in different aspects, it is necessary to design customized loss terms to guide the reconstruction of each component.
To be specific, the loss constraints imposed on the reflectance reconstruction include a structure loss Lstr and a color loss Lcolor, while the loss constraints imposed on the illumination reconstruction include a mean square error (MSE) loss Lmse.
The learned reflectance feature representations will focus more on the structural and color information due to the joint guidance of Lstr and Lcolor, while the learned illumination feature representations will focus more on the luminance information with the guidance of Lmse.
In the following, we will introduce the definitions and formulations of these loss terms one by one.
以下、これらの損失項の定義と定式化を1つずつ紹介する。
0.58
1) Structure loss: Previous works have reproted that the HVS is highly sensitive to the structural information of images and low-quality NTIs will change the structural percpetion [43].
1) 構造損失: HVS は画像の構造情報に非常に敏感であり,低品質の NTI が構造的摂動を変えることが示唆されている[43]。
0.84
We adopt the widely-used structural similarity (SSIM) [43] loss between the input reflectance image R and its corresponding reconstructed version ˆR for encouraging the encoder to have the capacity of extracting informative structural features.
Thus, a simple yet effective color loss term between R and ˆR is desired, which will encourage the encoder to have the capability of extracting color features.
3) MSE loss: For the reconstruction of illumination component, we only apply a simple MSE loss which is defined by the Euclidean distance between the L and ˆL:
3) MSE 損失:照明成分の再構成には、L と L の間のユークリッド距離によって定義される単純な MSE 損失のみを適用する。
(13) Constrained by the above loss terms, the content-related features and illumination-related features can be well extracted from the reflectance and illumination component, respectively.
We consider bilinear techniques to combine the reflectance and illumination feature representations into an unified one.
我々は,反射率と照明特徴表現を統一表現に結合するバイリニア手法を考える。
0.68
Bilinear models have shown powerful capability in modeling two-factor variations, such as style and content of images [45], location and appearance for fine-grained recognition [46], temporal and spatial aspects for video analysis [47], etc.
It also has been applied to address the BIQA problem where the synthetic and authentic distortions are modeled as the two-factor variations [48].
また、合成歪みと真正歪みを2要素変動[48]としてモデル化するBIQA問題にも適用されている。
0.70
Here, we tackle the blind NTIQE problem with a similar philosophy, where the reflectance and illumination components are modeled as the two-factor variations.
Given an input NTI and its side output feature maps from the reflectance and illumination encoders, C R are both with the size of hi × wi × di since the reflectance i and illumination encoder share the same architectures and configurations.
入力ntiとそのサイド出力特徴マップが反射および照明エンコーダから与えられると、crは、反射iと照明エンコーダが同じアーキテクチャと構成を共有しているため、どちらもhi × wi × diの大きさである。
0.83
Before performing bilinear pooling, C R and i are separately fed into a 1 × 1 conv layer to obtain their i C L corresponding compact version with 32 channels ( ˆC R and i ), i.e., hi × wi × 32.
バイリニアプーリングを行う前に、c r と i は別々に 1 × 1 conv 層に供給され、その ic l に対応するコンパクト版は 32 チャンネル ( \c r と i )、すなわち hi × wi × 32 となる。 訳抜け防止モード: バイリニアプーリングを行う前に、C R と i は分離して 1 × 1 の conv 層に供給され、32 チャンネル ( >C R と i ) のコンパクトバージョンに対応する i C L を得る。 i.e., hi × wi × 32。
0.81
Then, bilinear pooling is performed i ˆC L on C R
次に、c r 上の二線型プーリング i 〜c l を行う。
0.48
and C L i i and C L
cl iは i と C の L
0.67
i as follows:
私は下記の通りです。
0.43
Bi = ( ˆC R
bi = (\c r ) である。
0.55
i )T ˆC L i ,
i)T-C L i,
0.43
(14) where the outer product Bi is a vector of dimension 32 × 32.
(14) ここで、外積 Bi は次元 32 × 32 のベクトルである。
0.58
According to [49], bilinear representation is usually mapped from Riemannian manifold into an Euclidean space using signed square root and (cid:96)2 normalization [50]:
2 , (17) where Qk is the ground truth subjective quality score of the kth image in a mini-batch and ˆQk is the predicted quality score by DDB-Net.
2 , (17) qkはミニバッチにおけるkth画像の主観的品質スコアであり、qkはddb-netによる予測品質スコアである。 訳抜け防止モード: 2 , (17) qk is the ground truth subjective quality score of the kth image in a mini - batch and sqk ddb - netによって予測される品質スコアです。
0.52
It is noteworthy that bilinear pooling is a global strategy and therefore our DDB-Net can receive the input image with an arbitrary size.
We run 100 epoches with a learning rate decaying in the interval [3× 10−3, 3× 10−4].
学習速度は[3× 10−3, 3× 10−4]間隔で減衰する。
0.63
All the input images are resized into 512× 512× 3 before feeding into the network.
すべての入力画像はネットワークに送信する前に512×512×3にリサイズされる。
0.74
For the training of the whole DDB-Net, we also adopt the Adam with a learning rate of 3 × 10−5 for the target NTI database and use Batch normalization to stabilize the training process.
IV. EXPERIMENTAL RESULTS In this section, we first describe the experimental setups, including benchmark databases, evaluation protocols, and performance criteria.
Then, we compare the performance of DDBNet with state-of-the-art BIQA models on each individual database.
次に,DDBNetの性能を各データベース上の最先端BIQAモデルと比較する。
0.74
Finally, we conduct several ablation studies to justify the rationality of each critical component involved in DDBNet and present an application test by applying the DDBNet to automatic parameter tuning of an off-the-shelf NTI enhancement algorithm.
1) Benchmark Databases: The main experiments are conducted on the large-scale natural night-time image database (NNID) [25] which contains 2, 240 NTIs with 448 different image contents captured by three different photographic equipments (i.e., a digital camera (Device I: Nikon D5300), a mobile phone (Device II: iPhone 8plus) and a tablet (Device III: iPad mini2)) in real-world night-time scenarios.
The five settings are different for different image contents.
5つの設定は、画像の内容によって異なる。
0.72
In NNID, 1, 400 images with 280 different image contents are captured by Nikon D5300, 640 images with 128 different image contents are captured by iPhone 8plus, and 200 images with 40 different image contents are captured by iPad mini2.
Different from the NNID database which only contains raw NTIs, the EHND database contains both raw NTIs and their corresponding enhanced versions by different NTI enhancement algorithms.
7 2) Evaluation Protocols and Performance Criteria: We conduct experiments by following the general evaluation protocol adopted in existing learning-based BIQA studies.
For each time, we compute four criteria to measure the model performance.
毎回、モデル性能を測定するための4つの基準を計算します。
0.70
The four performance criteria include Pearson linear correlation coefficient (PLCC), Spearman rank order correlation coefficient (SRCC), Kendall rank order correlation coefficient (KRCC), and root mean square error (RMSE).
4つの性能基準は,Pearson linear correlation coefficient (PLCC), Spearman rank order correlation coefficient (SRCC), Kendall rank order correlation coefficient (KRCC), root mean square error (RMSE)である。
0.79
Among these criteria, PLCC and RMSE measure the prediction precision while SRCC and KRCC measure the prediction monotonicity.
These criteria results from the five sessions are calculated respectively and averaged to serve as the final model performance.
これら5セッションの基準値をそれぞれ算出し、最終モデルのパフォーマンスとして平均化する。
0.74
B. Performance Comparisons Since there is always no available pristine reference for real-world NTIs, the quality evaluation of NTIs can only be performed in a no-reference/blind manner.
The trainingbased ones commonly adopt elaborately designed features to characterize the level of deviations from statistical regularities of natural scenes, based on which a quality prediction function is learned via supprot vector regression (SVR) [53].
The training-free ones first build a pristine statistical model from a large collection of high-quality natural images and then measure the distance between this pristine statistical model and the statistical model of the distorted image as the estimated quality score.
By contrast, the deep learning-based BIQA methods directly optimize an end-to-end function mapping from the input image to its quality score while without any effort on manual feature engineering.
1) Comparisons on NNID: The performance comparison results of different BIQA methods on the NNID database are shown in Table I. From the results, we can have the following observations.
First, most training-based BIQA models perform better than the two training-free ones (i.e., NIQE and ILNIQE) while the two deep learning-based BIQA models (i.e., WaDIQaM and DBCNN) are superior to most handcraft feature-based BIQA methods.
It is reasonable because BIQA is a challenging task where training is particularly useful to model the complex non-linear relationship between the extracted features and perceived quality score, and end-to-end deep learning technique further provides an effective solution image-to-quality mapping to directly establish the explicit
function owing to its powerful capacity of automatic feature representation learning.
自動特徴表現学習の強力な能力によって機能する。
0.85
Second, the existing NSS featurebased BIQA cannot obtain satisfactory results for evaluating NTIs as NSS is not that suitable to characterize the degradation
Third, the proposed DDBNet delivers the best performance among all the competitors in terms of all performance criteria, i.e., the highest PLCC, SRCC, KRCC values and the lowest RMSE value.
In the figure, 1/-1 indicates that row models perform statistically better/worse than the column models.
図1/1は、列モデルがカラムモデルよりも統計的に良い/低いパフォーマンスを示す。
0.75
the most performance results of different BIQA methods on the EHND database are shown in Table II and the significance t-test results are shown in Fig 8.
It is observed from these results that our proposed DDB-Net outperforms other competitors by a large margin in terms of all performance criteria on the EHND database.
In this case, important role of NTIQE is to automatically select the one with the highest visual quality from 15 enhanced results generated from the same NTI.
Specifically, we measure the rank-n accuracy which is closely relevant with the capability of a certain objective quality metric in selecting the optimal enhanced result from a set of candidates.
Given 15 different enhanced results associated with the same raw NTI, the rank-n accuracy is defined as the percentage of images whose top-1 result in terms of MOS
In the figure, 1/-1 indicates that row models perform statistically better/worse than the column models.
図1/1は、列モデルがカラムモデルよりも統計的に良い/低いパフォーマンスを示す。
0.75
is that we have dedicated decomposing the sophisticated blind NTIQE task into two easier sub-tasks with each sub-task accounting for illumination perception and content perception, respectively.
In such a way, the features related to the illumination perception and the content perception can be better learned and finally fused to facilitate blind NTIQE.
In addition to the numerical performance results, we also show the scatter plots between the objective scores (predicted by BIQA methods) and the subjective MOSs (provided in the database) in Fig 6.
In the scatter plot, each point corresponds to an image in the NNID database and the x-axis represents the prediction scores by BIQA methods while the y-axis represents the ground truth subjective MOSs.
2) Comparisons on EHND: A well-performing NTIQE should also be able to measure the performance of different NTI quality enhancement algorithms, i.e., well evaluate different enhanced results.
Actually, a certain enhancement algorithm may result in particularly bad enhanced result which may still suffer from unsatisfactory brightness and even more serious color distortions than the original raw NTI.
Therefore, we also evaluate the performance of different BIQA methods on another nighttime image database EHND which contains 1, 500 images obtained by applying 15 off-the-shelf NTI enhancement algorithms on 100 raw NTIs.
It is observed that our DDB-Net always delivers highest rank-n accuracy values, indicating the best capability in selecting the one with the highest visual quality from a set of candidates.
ddb-netは、常に最もランクnの精度の高い値を提供しており、候補群から最も視覚品質の高いものを選択するのに最適な能力を示しています。 訳抜け防止モード: DDB - Netは常に最高ランク - n の精度の値を提供しています。 候補者の中から最高の視覚的品質の人を選ぶ最高の能力を示す。
0.73
C. Application: Automatic Parameter Tuning of NTI Quality Enhancement Algorithm
C. 応用:NTI品質向上アルゴリズムの自動パラメータ調整
0.84
An effective blind NTIQE should be able to well guide the optimization of NTI quality enhancement algorithms.
効果的なブラインドNTIQEは、NTI品質向上アルゴリズムの最適化をうまく導くことができる。
0.79
In this section, we demonstrate this idea by applying the proposed DDB-Net to automatic parameter tuning of off-the-shelf NTI quality enhancement algorithms.
It is challenging and time-consuming to handpick a set of parameters that work well for all image contents.
すべての画像コンテンツでうまく機能するパラメータのセットを手で選択するのは困難で時間がかかります。
0.69
A well-performing blind NTIQE is able to replace the role of humans in this task, especially when the volume of images to be processed is particularly large.
Here, we use the LIME algorithm [55] as a representative example of NTI quality enhancement algorithm, which involves two tunable parameters g and l.
ここでは、2つの可変パラメータ g と l を含む NTI 品質向上アルゴリズムの代表的な例として LIME アルゴリズム [55] を用いる。
0.89
The default values are: g = 0.6 and l = 0.2.
デフォルト値は g = 0.6 と l = 0.2 である。
0.87
However, the visual quality of the final enhanced image is highly sensitive to these two parameters.
しかし、最終的な強調画像の視覚品質は、これら2つのパラメータに対して非常に敏感である。
0.64
Fig 10 shows example images generated with different g and l values.
fig 10は、gとlの異なる値で生成された例を示す。
0.64
In the figure, warmer color indicates better predicted quality of the corresponding enhanced image.
図中では、より暖かい色は、対応する強調画像の予測品質の向上を示す。
0.69
The corresponding scores predicted by our DDB-Net are also shown under each image.
ddb-netによって予測される対応するスコアも各画像の下に表示される。
0.53
By varying g and l, we can obtain enhanced results with significantly different visual quality.
g と l の変化により、視覚的品質が著しく異なる拡張結果が得られる。
0.76
For example, the two enhanced results in the left side of Fig 10 still suffers from over-/under exposure problem while the two enhanced results in the right side exhibits much better visual quality with much more finer details and natural color appearance.
It is found that our DDB-Net can evaluate their visual qualities consistently with human subjective perception.
ddb-netは人間の主観的知覚と一貫してその視覚特性を評価できることが判明した。
0.60
Furthermore, we also find that the visual quality of the upper right image is better than that of the bottom right one which is
さらに、右上像の視覚的品質が右下像の視覚的品質よりも優れていることも判明した。
0.65
produced by using the default parameter values.
デフォルトパラメータ値を使用して生成される。
0.74
It means that it is possible to adaptively determine the optimal parameter values under the guidance of our proposed DDN-Net.
これは,提案するddn-netの指導の下で最適パラメータ値を適応的に決定できることを意味する。
0.74
V. CONCLUSION
V.コンキュレーション
0.76
This paper has presented a novel deep NTIQE called DDBNet which consists of three modules namely image decomposition module, feature encoding module, and bilinear pooling module.
With the help of decomposing the input NTI into two independent layer components (illumination and reflectance), the degradation features related to illumination perception and content perception are better learned and then fused with bilinear pooling to improve the performance of blind NTIQE.
[2] Y. Zhan and R. Zhang, “No-reference jpeg image quality assessment based on blockiness and luminance change,” IEEE Signal Processing Letters, vol.
ieee signal processing letters, vol. “no-reference jpeg image quality assessment based on blockiness and luminance change”. [2] y. zhanとr. zhang。 訳抜け防止モード: [2 ]Y. Zhan, R. Zhang, “No- Reference jpeg Image Quality Assessment based on blockiness and luminance change” IEEE Signal Processing Letters , vol .
0.45
24, no. 6, pp. 760–764, 2017.
24 no. 6, pp. 760-764, 2017 頁。
0.83
[3] S. A. Golestaneh and D. M. Chandler, “No-reference quality assessment of jpeg images via a quality relevance map,” IEEE Signal Processing Letters, vol.
s. a. golestanehとd. m. chandlerは、“品質関連マップによるjpeg画像の参照品質評価は行わない”、とieee signal processing letters, vol. は述べている。 訳抜け防止モード: [3 ]S.A.ゴレスタインとD.M.チャンドラーは「品質関連マップによるjpeg画像の基準品質評価」と評した。 IEEE Signal Processing Letters , vol .
0.80
21, no. 2, pp. 155–158, 2014.
21 no. 2, pp. 155–158, 2014年。
0.91
[4] K. Gu, G. Zhai, W. Lin, X. Yang, and W. Zhang, “No-reference image sharpness assessment in autoregressive parameter space,” IEEE Transactions on Image Processing, vol.
K. Gu, G. Zhai, W. Lin, X. Yang, W. Zhang, “No-reference image sharpness Assessment in autoregressive parameters space”, IEEE Transactions on Image Processing, vol。 訳抜け防止モード: 4 ] k. gu, g. zhai, w. lin, x. yang, and w. zhang, "no - reference image sharpness assessment in autoregressive parameter space" ieeeは画像処理でトランザクションを処理します。
0.75
24, no. 10, pp. 3218–3231, 2015.
24 no. 10, pp. 3218-3231, 2015年。
0.87
[5] T. Oh, J. Park, K. Seshadrinathan, S. Lee, and A. C. Bovik, “No-reference sharpness assessment of camera-shaken images by analysis of spectral structure,” IEEE Transactions on Image Processing, vol.
5] t. oh, j. park, k. seshadrinathan, s. lee, a. c. bovik, “スペクトル構造の分析によるカメラシェーク画像の非参照シャープネス評価”。 訳抜け防止モード: 5 ] t. オ・j・パーク k・セシャドリナサン s. lee, and a. c. bovik, "no - reference sharpness assessment of camera - shaken images by analysis of spectral structure" ieeeは画像処理でトランザクションを処理します。
0.76
23, no. 12, pp. 5428–5439, 2014.
23, No. 12, pp. 5428–5439, 2014。
0.47
lgQQ= 0.8974Q= 0.6347Q= 0.6732Q= 0.8335
lgQ= 0.8974Q= 0.6347Q= 0.6732Q= 0.8335
0.16
英語（論文から抽出）
日本語訳
スコア
[6] C. Tang, X. Yang, and G. Zhai, “Noise estimation of natural images via statistical analysis and noise injection,” IEEE Transactions on Circuits and Systems for Video Technology, vol.
IEEE Transactions on Circuits and Systems for Video Technology,vol.[6]C. Tang, X. Yang, G. Zhai, “Noise Estimation of natural image via statistics analysis and noise Injection”. IEEE Transactions on Circuits and Systems for Video Technology. 訳抜け防止モード: [6]c.唐、x.yang、g.zhai。 統計解析とノイズインジェクションによる自然画像の雑音推定 ieee transactions on circuits and systems for video technology, vol。
0.69
25, no. 8, pp. 1283–1294, 2015.
25 no. 8, pp. 1283-1294, 2015 頁。
0.82
[7] M. A. Saad, A. C. Bovik, and C. Charrier, “Blind image quality assessment: A natural scene statistics approach in the DCT domain,” IEEE Transactions on Image Processing, vol.
M.A. Saad, A. C. Bovik, C. Charrier, “Blind Image Quality Assessment: A natural scene statistics approach in the DCT domain”, IEEE Transactions on Image Processing, Vol. 訳抜け防止モード: 【7 ]m.a.サード,a.c.ボヴィク, c. charrier, “ blind image quality assessment: a natural scene statistics approach in the dct domain” ieeeは画像処理でトランザクションを処理します。
0.75
21, no. 8, pp. 3339–3352, 2012.
21 no. 8, pp. 3339-3352, 2012 頁。
0.78
[8] A. Mittal, A. K. Moorthy, and A. C. Bovik, “No-reference image quality assessment in the spatial domain,” IEEE Transactions on Image Processing, vol.
A. Mittal, A. K. Moorthy, A. C. Bovik, “No-reference image quality Assessment in the spatial domain, IEEE Transactions on Image Processing, vol. 訳抜け防止モード: [8 ] A. Mittal, A. K. Moorthy, A. C. Bovik, 「空間領域における基準画像品質評価」 IEEE Transactions on Image Processing , vol。
0.85
21, no. 12, pp. 4695–4708, 2012.
第21巻第12号, pp. 4695-4708, 2012。
0.56
[9] L. Liu, H. Dong, H. Huang, and A. C. Bovik, “No-reference image quality assessment in curvelet domain,” Signal Processing: Image Communication, vol.
9] L. Liu, H. Dong, H. Huang, A. C. Bovik, “No-reference image quality Assessment in curvelet domain”, Signal Processing: Image Communication, vol。 訳抜け防止モード: 9 ] l. liu, h. dong, h. huang, a.c. bovik氏, “no - reference image quality assessment in curvelet domain” 信号処理 : 画像通信, vol.
0.83
29, no. 4, pp. 494–505, 2014.
29, No. 4, pp. 494–505, 2014。
0.47
[10] A. K. Moorthy and A. C. Bovik, “Blind image quality assessment: From natural scene statistics to perceptual quality,” IEEE Transactions on Image Processing, vol.
A. K. Moorthy, A. C. Bovik, “Blind Image Quality Assessment: From natural scene statistics to perceptual quality”, IEEE Transactions on Image Processing, vol. 訳抜け防止モード: 10 ] a.k.moorthy と a.c. bovik の「盲眼画像品質評価 : 自然景観統計から知覚品質まで」 ieeeは画像処理でトランザクションを処理します。
0.74
20, no. 12, pp. 3350– 3364, 2011.
20 no. 12 pp. 3350–3364, 2011年。
0.78
[11] Q. Li, W. Lin, J. Xu, and Y. Fang, “Blind image quality assessment using statistical structural and luminance features,” IEEE Transactions on Multimedia, vol.
IEEE Transactions on Multimedia, vol. [11] Q. Li, W. Lin, J. Xu, Y. Fang, “Blind image quality Assessment using statistics structure and luminance features, IEEE Transactions on Multimedia, vol. 訳抜け防止モード: [11]Q.Li,W.Lin,J.Xu, そしてY.Fang氏は,“統計的構造と輝度特徴を用いたブラインド画像の品質評価”を行った。 IEEE Transactions on Multimedia , vol。
0.71
18, no. 12, pp. 2457– 2469, 2016.
18, no. 12 pp. 2457–2469, 2016年。
0.84
[12] K. Gu, G. Zhai, X. Yang, and W. Zhang, “Using free energy principle for blind image quality assessment,” IEEE Transactions on Multimedia, vol.
IEEE Transactions on Multimedia, vol.[12] K. Gu, G. Zhai, X. Yang, W. Zhang, “Free Energy principle for blind image quality Assessment”. IEEE Transactions on Multimedia. 訳抜け防止モード: [12]K.Gu,G.Zhai,X.Yang, W. Zhangは曰く、“自由エネルギーの原則を使って画像のクオリティを評価する”。 IEEE Transactions on Multimedia , vol。
0.75
17, no. 1, pp. 50–63, 2015.
17, No. 1, pp. 50-63, 2015。
0.91
[13] W. Xue, X. Mou, L. Zhang, A. C. Bovik, and X. Feng, “Blind image quality assessment using joint statistics of gradient magnitude and Laplacian features,” IEEE Transactions on Image Processing, vol.
13] w. xue, x. mou, l. zhang, a. c. bovik, x. feng, “グラデーションサイズとラプラシアン特徴の合同統計を用いた画像品質アセスメント”。 訳抜け防止モード: [13 ]W. Xue, X. Mou, L. Zhang, A. C. Bovik, X. Fengは,「勾配等級とラプラシアン特徴の合同統計を用いたブラインド画像品質評価」と評した。 IEEE Transactions on Image Processing , vol。
0.82
23, no. 11, pp. 4850–4862, 2014.
23 no. 11, pp. 4850–4862, 2014 頁。
0.87
[14] Q. Li, W. Lin, and Y. Fang, “No-reference quality assessment for multiply-distorted images in gradient domain,” IEEE Signal Processing Letters, vol.
Q. Li, W. Lin, Y. Fang, “No-reference quality Assessment for multiply-distorted images in gradient domain”, IEEE Signal Processing Letters, vol。 訳抜け防止モード: 14 ] q. li, w. lin, y. fang. 「勾配領域における多重歪画像に対する基準品質評価」 ieee 信号処理文字 , vol.
0.75
23, no. 4, pp. 541–545, 2016.
23 no. 4, pp. 541–545, 2016年。
0.90
[15] L. Liu, B. Liu, H. Huang, and A. C. Bovik, “No-reference image quality assessment based on spatial and spectral entropies,” Signal Processing: Image Communication, vol.
L. Liu, B. Liu, H. Huang, A. C. Bovik, “No-reference image quality Assessment based on space and spectrum entropies”, Signal Processing: Image Communication, vol。 訳抜け防止モード: [15 ]L.Lu,B.Lu,H.Huang, そしてA・C・ボヴィクは「空間エントロピーとスペクトルエントロピーに基づく基準画像品質評価」と評した。 信号処理 : 画像通信, vol。
0.68
29, no. 8, pp. 856–863, 2014.
29, 8, pp. 856-863, 2014 頁。
0.72
[16] K. Gu, D. Tao, J.
[16]K. Gu, D. Tao, J.
0.49
-F. Qiao, and W. Lin, “Learning a no-reference quality assessment model of enhanced images with big data,” IEEE Transactions on Neural Networks and Learning Systems, vol.
-F。 qiao, and w. lin, “ビッグデータを用いた強化画像の参照なし品質評価モデルを学ぶ”と、ieee transactions on neural networks and learning systems, vol. は述べている。
0.58
29, no. 4, pp. 1301–1313, 2017.
29, No. 4, pp. 1301–1313, 2017。
0.45
[17] L. Zhang, L. Zhang, and A. C. Bovik, “A feature-enriched completely blind image quality evaluator,” IEEE Transactions on Image Processing, vol.
A feature-enriched completely blind image quality evaluator, IEEE Transactions on Image Processing, vol.[17] L. Zhang, L. Zhang, A. C. Bovik。 訳抜け防止モード: [17 ] L. Zhang, L. Zhang, A. C. Bovik IEEE Transactions on Image Processing,vol.“機能 – 完全に盲目な画像品質評価器の強化”。
0.84
24, no. 8, pp. 2579–2591, 2015.
24, No. 8, pp. 2579–2591, 2015。
0.95
[18] A. Mittal, R. Soundararajan, and A. C. Bovik, “Making a “completely blind” image quality analyzer,” IEEE Signal Processing Letters, vol.
[18] a. mittal, r. soundararajan, a. c. bovik, “making a completely blind” image quality analyzer”, ieee signal processing letters, vol.] (英語) 訳抜け防止モード: [18 ]A. Mittal, R. Soundararajan, A. C. Bovik 完全に盲目」な画像品質分析装置を作る」 IEEE Signal Processing Letters , vol .
0.81
20, no. 3, pp. 209–212, 2012.
20. 3, pp. 209-212, 2012 頁。
0.79
[19] Q. Li, W. Lin, and Y. Fang, “BSD: Blind image quality assessment based on structural degradation,” Neurocomputing, vol.
Q. Li, W. Lin, Y. Fang, “BSD: 構造劣化に基づくブラインド画像の品質評価”, Neurocomputing, vol。
0.70
236, pp. 93–103, 2017.
236, pp. 93-103, 2017。
0.42
[20] P. Ye, J. Kumar, L. Kang, and D. Doermann, “Unsupervised feature learning framework for no-reference image quality assessment,” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 1098–1105.
P. Ye, J. Kumar, L. Kang, D. Doermann, “Unsupervised feature learning framework for no-reference image quality Assessment” in 2012 IEEE Conference on Computer Vision and Pattern Recognition, 2012 pp. 1098–1105。 訳抜け防止モード: [20]P.Ye,J.Kumar,L.Kang, そしてD. Doermann氏は,“No-Reference Image Qualityアセスメントのための教師なし機能学習フレームワーク”だ。 2012年IEEE Conference on Computer Vision and Pattern Recognition, 2012, pp. 1098–1105.
0.81
[21] J. Xu, P. Ye, Q. Li, H. Du, Y. Liu, and D. Doermann, “Blind image quality assessment based on high order statistics aggregation,” IEEE Transactions on Image Processing, vol.
J.Xu, P. Ye, Q. Li, H. Du, Y. Liu, D. Doermann, “Blind image quality Assessment based on High order statistics aggregate, IEEE Transactions on Image Processing, vol. 訳抜け防止モード: [21 ]J.Xu,P.Ye,Q.Li, H. Du, Y. Liu, D. Doermann, “高次統計集計に基づくブラインド画像の品質評価”。 IEEE Transactions on Image Processing , vol。
0.81
25, no. 9, pp. 4444–4457, 2016.
25 no. 9, pp. 4444-4457, 2016年。
0.87
[22] Q. Jiang, F. Shao, G. Jiang, M. Yu, and Z. Peng, “Supervised dictionary learning for blind image quality assessment using quality-constraint sparse coding,” Journal of Visual Communication and Image Representation, 2015.
22] q. jiang, f. shao, g. jiang, m. yu, z. peng, “supervised dictionary learning for blind image quality assessment using quality-constraint sparse coding” journal of visual communication and image representation, 2015” (英語) 訳抜け防止モード: [22]Q.ジャン、F.シャオ、G.ジャン、 M. Yu, Z. Peng, “品質-制約スパースコーディングによる視覚的画像品質評価のための教師付き辞書学習”。 Journal of Visual Communication and Image Representation , 2015
0.77
[23] Q. Jiang, F. Shao, W. Lin, K. Gu, G. Jiang, and H. Sun, “Optimizing multistage discriminative dictionaries for blind image quality assessment,” IEEE Transactions on Multimedia, vol.
Q. Jiang, F. Shao, W. Lin, K. Gu, G. Jiang, H. Sun, “多段階の識別辞書をブラインド画像の品質評価に最適化する”IEEE Transactions on Multimedia, Vol.
0.85
20, no. 8, pp. 2035–2048, 2018.
20.8、p.2035-2048、2018。
0.51
[24] Q. Wu, H. Li, F. Meng, K. N. Ngan, B. Luo, C. Huang, and B. Zeng, “Blind image quality assessment based on multichannel feature fusion and label transfer,” IEEE Transactions on Circuits and Systems for Video Technology, vol.
[24] Q. Wu, H. Li, F. Meng, K. N. Ngan, B. Luo, C. Huang, B. Zeng, “Blind image quality Assessment based on multi channel feature fusion and label transfer, IEEE Transactions on Circuits and Systems for Video Technology, vol. 訳抜け防止モード: [24 ]Q. Wu, H. Li, F. Meng, K. N. Ngan, B. Luo, C. Huang, B. Zeng 「多チャンネル機能融合とラベル転送に基づくブラインド画像品質評価」 IEEE Transactions on Circuits and Systems for Video Technology, vol.
0.87
26, no. 3, pp. 11
26、p.3。 11
0.48
425–440, 2016.
425–440, 2016.
0.42
[25] T. Xiang, Y. Yang, and S. Guo, “Blind night-time image quality assessment: Subjective and objective approaches,” IEEE Transactions on Multimedia, vol.
[25]T.Xiang,Y. Yang,S. Guo, “Blind night-time image quality Assessment: Subjective and objective approach”, IEEE Transactions on Multimedia, vol。 訳抜け防止モード: [25 ]T.Xiang、Y.Y.Yang、S.Guo 「盲夜-時間画像品質評価 : 主観的・客観的アプローチ」 IEEE Transactions on Multimedia , vol。
0.68
22, no. 5, pp. 1259–1272, 2020.
22 no. 5, pp. 1259-1272, 2020。
0.82
[26] J. McCann, Retinex Theory.
J. McCann, Retinex Theory (英語)
0.67
New York, NY: Springer New
New York, NY: Springer New
0.42
York, 2016, pp. 1118–1125.
ヨーク、2016年、p.1118-1125。
0.53
[27] J. Wu, J. Zeng, Y. Liu, G. Shi, and W. Lin, “Hierarchical feature degradation based blind image quality assessment,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 510–517.
[27] j. wu, j. zeng, y. liu, g. shi, w. lin, “hierarchical feature degradation based blind image quality assessment” in the proceedings of the ieee international conference on computer vision, 2017 pp. 510–517. (英語) 訳抜け防止モード: [27 ]J.Wu,J.Zeng,Y.Lu, G. Shi, W. Lin, “階層的特徴劣化に基づくブラインド画像品質評価” In Proceedings of the IEEE International Conference on Computer Vision, 2017 510-517頁。
0.83
[28] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770–778.
[28] k. he, x. zhang, s. ren, and j. sun, “deep residual learning for image recognition” in the proceedings of the ieee conference on computer vision and pattern recognition, 2016 pp. 770–778. コンピュータビジョンとパターン認識に関する国際会議の議事録。
0.73
[29] Y. Li, L.
[29]Y,Li,L。
0.40
-M. Po, L. Feng, and F. Yuan, “No-reference image quality assessment with deep convolutional neural networks,” in IEEE International Conference on Digital Signal Processing, 2016, pp. 685–689.
-M。 Po, L. Feng, and F. Yuan, “No-reference image quality Assessment with Deep Convolutional Neural Network” in IEEE International Conference on Digital Signal Processing, 2016, pp. 685–689。 訳抜け防止モード: -M。 と、Po, L. Feng, F. Yuan氏は述べている。 IEEE International Conference on Digital Signal Processing, 2016, pp. 685–689.
0.48
[30] J. Deng, W. Dong, R. Socher, L.
30] j. deng, w. dong, r. socher, l.
0.39
-J. Li, K. Li, and L. FeiFei, “Imagenet: A large-scale hierarchical image database,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009, pp. 248–255.
-j。 Li, K. Li, and L. FeiFei, “Imagenet: A Large-scale hierarchical image database” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2009 pp. 248–255。
0.59
[31] L. Kang, P. Ye, Y. Li, and D. Doermann, “Convolutional neural networks for no-reference image quality assessment,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2014, pp. 1733–1740.
L.Kang, P. Ye, Y. Li, D. Doermann, “Convolutional Neural Network for no-reference image quality Assessment” in Proceedings on the IEEE Conference on Computer Vision and Pattern Recognition, 2014 pp. 1733–1740。 訳抜け防止モード: [31 ]L.Kang,P.Y,Y.Li, そしてD. Doermannは、”畳み込みニューラルネットワークのノー参照画像品質評価”について語る。 IEEE Conference on Computer Vision and Pattern Recognition に参加して 2014 , pp . 1733–1740 .
0.82
[32] Y. Li, L.
[32]Y,Li,L。
0.41
-M. Po, X. Xu, L. Feng, F. Yuan, C.
-M。 Po, X. Xu, L. Feng, F. Yuan, C。
0.44
-H. Cheung, and K.
-h。 チャン、K。
0.61
-W. Cheung, “No-reference image quality assessment with shearlet transform and deep neural networks,” Neurocomputing, vol.
-W。 cheung, “no-reference image quality assessment with shearlet transform and deep neural networks”, neurocomputing, vol。
0.61
154, pp. 94–109, 2015.
154, pp. 94-109, 2015。
0.88
[33] X. Liu, J. van de Weijer, and A. D. Bagdanov, “Rankiqa: Learning from rankings for no-reference image quality assessment,” in Proceedings of the IEEE International Conference on Computer Vision, 2017, pp. 1040–1049.
X. Liu, J. van de Weijer, A. D. Bagdanov, “Rankiqa: Learning from rankings for no-reference image quality Assessment” in Proceedings of the IEEE International Conference on Computer Vision, 2017 pp. 1040–1049。
0.42
[34] K. Ma, W. Liu, K. Zhang, Z. Duanmu, Z. Wang, and W. Zuo, “End-to-end blind image quality assessment using deep neural networks,” IEEE Transactions on Image Processing, vol.
K. Ma, W. Liu, K. Zhang, Z. Duanmu, Z. Wang, W. Zuo, “End-to-end blind image quality Assessment using Deep Neural Network”, IEEE Transactions on Image Processing, vol。 訳抜け防止モード: [34 ] K. Ma, W. Liu, K. Zhang, Z. Duanmu, Z. Wang, W. Zuo氏は,“ディープニューラルネットワークによる視覚的品質評価の終端 - 終端 - 終端 - 終端)”と述べている。 IEEE Transactions on Image Processing , vol。
0.65
27, no. 3, pp. 1202–1213, 2017.
27 no. 3, pp. 1202-1213, 2017 頁。
0.41
[35] S. Bosse, D. Maniry, K.
[35]S. Bosse, D. Maniry, K.
0.48
-R. M¨uller, T. Wiegand, and W. Samek, “Deep neural networks for no-reference and full-reference image quality assessment,” IEEE Transactions on Image Processing, vol.
[36] W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “Blind image quality assessment using a deep bilinear convolutional neural network,” IEEE Transactions on Circuits and Systems for Video Technology, vol.
[36] w. zhang, k. ma, j. yan, d. deng, z. wang, “ディープバイリニア畳み込みニューラルネットワークを用いた画像品質評価” ieee transactions on circuits and systems for video technology, vol.)。 訳抜け防止モード: [36 ] W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “深層双線形畳み込みニューラルネットワークによるブラインド画像の品質評価”。 IEEE Transactions on Circuits and Systems for Video Technology, vol.
0.90
30, no. 1, pp. 36–47, 2018.
30 no. 1 pp. 36-47、2018年。
0.76
[37] X. Min, G. Zhai, K. Gu, X. Yang, and X. Guan, “Objective quality evaluation of dehazed images,” IEEE Transactions on Intelligent Transportation Systems, vol.
[37] X. Min, G. Zhai, K. Gu, X. Yang, X. Guan, “Objective quality evaluation of dehazed images, IEEE Transactions on Intelligent Transportation Systems, vol.
0.46
20, no. 8, pp. 2879– 2892, 2018.
20, 8, pp. 2879–2892, 2018。
0.72
[38] X. Min, G. Zhai, K. Gu, Y. Zhu, J. Zhou, G. Guo, X. Yang, X. Guan, and W. Zhang, “Quality evaluation of image dehazing methods using synthetic hazy images,” IEEE Transactions on Multimedia, vol.
[38] X. Min, G. Zhai, K. Gu, Y. Zhu, J. Zhou, G. Guo, X. Yang, X. Guan, W. Zhang, “Quality Evaluation of image dehazing method using synthetic hazy image, IEEE Transactions on Multimedia, vol. 訳抜け防止モード: [38 ]X. Min, G. Zhai, K. Gu, Y. Zhu, J. Zhou, G. Guo, X. Yang X. Guan, W. Zhang, “Synthetic hazy imageを用いた画像脱ヘイズ手法の品質評価” IEEE Transactions on Multimedia , vol。
0.92
21, no. 9, pp. 2319–2333, 2019.
第21巻第9号, pp. 2319–2333, 2019。
0.70
[39] Q. Wu, L. Wang, K. N. Ngan, H. Li, and F. Meng, “Beyond synthetic data: A blind deraining quality assessment metric towards authentic rain image,” in IEEE International Conference on Image Processing, 2019, pp. 2364–2368.
IEEE International Conference on Image Processing, 2019, pp. 2364–2368.[39] Q. Wu, L. Wang, K. N. Ngan, H. Li, F. Meng, “Beyond synthetic data: A blind deraining quality Assessment metric to a secure rain image”. IEEE International Conference on Image Processing, 2019, pp. 2364–2368。 訳抜け防止モード: [39 ]Q.Wu,L.Wang,K.N.Ngan , H. Li, F. Meng, “Beyond synthetic data : A blind deraining quality Assessment metric to a Agentic rain image” IEEE International Conference on Image Processing, 2019, pp. 2364–2368
0.43
[40] Q. Wu, L. Wang, K. N. Ngan, H. Li, F. Meng, and L. Xu, “Subjective and objective de-raining quality assessment towards authentic rain image,” IEEE Transactions on Circuits and Systems for Video Technology, vol.
40] q. wu, l. wang, k. n. ngan, h. li, f. meng, l. xu, “本物の雨画像に対する主観的かつ客観的な降雨品質評価”。
0.52
30, no. 11, pp. 3883–3897, 2020.
30 no. 11, pp. 3883-3897, 2020。
0.82
[41] M. Yang and A. Sowmya, “An underwater color image quality evaluation metric,” IEEE Transactions on Image Processing, vol.
IEEE Transactions on Image Processing, vol.[41]M. Yang, A. Sowmya, “An underwater color image quality evaluation metric, IEEE Transactions on Image Processing, Vol. 訳抜け防止モード: [41 ]m. yang, a. sowmya, “an underwater color image quality evaluation metric” ieeeは画像処理でトランザクションを処理します。
0.73
24, no. 12, pp. 6062–6071, 2015.
24巻12頁、p.6062-6071、2015年。
0.59
[42] P. Guo, L. He, S. Liu, D. Zeng, and H. Liu, “Underwater image quality assessment: Subjective and objective methods,” IEEE Transactions on Multimedia, 2021.
[42]P. Guo, L. He, S. Liu, D. Zeng, H. Liu, “Underwater image quality Assessment: Subjective and objective Method”, IEEE Transactions on Multimedia, 2021。 訳抜け防止モード: [42 ]P.Guo, L. He, S. Liu, D. Zeng, H. Liu, 「水中画像品質評価 : 主観的・客観的手法」 IEEE Transactions on Multimedia , 2021
0.81
英語（論文から抽出）
日本語訳
スコア
12 [43] Z. Wang, A. Bovik, H. Sheikh, and E. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Transactions on Image Processing, vol.
12 Z.Wang, A. Bovik, H. Sheikh, E. Simoncelli, “画像品質評価:エラーの可視性から構造的類似性に至るまで”, IEEE Transactions on Image Processing, vol。
0.63
13, no. 4, pp. 600–612, 2004.
13, No. 4, pp. 600-612, 2004。
0.91
[44] A. Ignatov, N. Kobyshev, R. Timofte, and K. Vanhoey, “Dslrquality photos on mobile devices with deep convolutional networks,” in 2017 IEEE International Conference on Computer Vision (ICCV), 2017, pp. 3297–3305.
A. Ignatov, N. Kobyshev, R. Timofte, K. Vanhoey, “Dslrquality photos on mobile devices with deep convolutional network” in 2017 IEEE International Conference on Computer Vision (ICCV), 2017 pp. 3297–3305。 訳抜け防止モード: 44] a. ignatov, n. kobyshev, r. timofte, k. vanhoey. 2017年ieee international conference on computer vision (iccv) で "dslrquality photos on mobile devices with deep convolutional networks, ” と題された。 2017 , pp . 3297–3305 .
0.85
[45] J. B. Tenenbaum and W. T. Freeman, “Separating style and
[45]J・B・テネンバウムとW・T・フリーマン「分離様式」
0.58
content,” in NIPS, 1996.
とnips, 1996に書いている。
0.56
[46] T. -Y.
[46]t。 -y。
0.34
Lin, A. RoyChowdhury, and S. Maji, “Bilinear cnn models for fine-grained visual recognition,” 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1449–1457, 2015.
Lin, A. RoyChowdhury, and S. Maji, “Bilinear cnn models for fine-fine visual recognition”. 2015 IEEE International Conference on Computer Vision (ICCV), pp. 1449–1457, 2015.
0.48
[47] K. Simonyan and A. Zisserman, “Two-stream convolutional
[47]K. SimonyanとA. Zisserman, “2ストリームの畳み込み”
0.75
networks for action recognition in videos,” in NIPS, 2014.
とNIPS, 2014で述べている。
0.34
[48] W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “Blind image quality assessment using a deep bilinear convolutional neural network,” IEEE Transactions on Circuits and Systems for Video Technology, vol.
[48] w. zhang, k. ma, j. yan, d. deng, z. wang, “ディープバイリニア畳み込みニューラルネットワークを用いた画像品質評価” ieee transactions on circuits and systems for video technology, vol.)。 訳抜け防止モード: [48 ] W. Zhang, K. Ma, J. Yan, D. Deng, and Z. Wang, “深層双線形畳み込みニューラルネットワークによるブラインド画像の品質評価”。 IEEE Transactions on Circuits and Systems for Video Technology, vol.
0.89
30, pp. 36–47, 2020.
第30巻、p.36-47、2020。
0.42
[49] X. Pennec, P. Fillard, and N. Ayache, “A riemannian framework for tensor computing,” International Journal of Computer Vision, vol.
[49] x. pennec, p. fillard, and n. ayache, “a riemannian framework for tensor computing”, international journal of computer vision, vol. (英語) 訳抜け防止モード: [49 ]X. Pennec, P. Fillard, N. Ayache テンソルコンピューティングのためのリーマン的フレームワーク” International Journal of Computer Vision,vol
0.68
66, pp. 41–66, 2005.
66, pp. 41-66, 2005。
0.86
[50] F. Perronnin, J. S´anchez, and T. Mensink, “Improving the fisher
50] f. perronnin, j. s ́anchez, t. mensink, "improving the fisher"
0.42
kernel for large-scale image classification,” in ECCV, 2010.
大規模な画像分類のためのカーネル” ECCV, 2010。
0.85
[51] D. P. Kingma and J. Ba, “Adam: A method for stochastic
[51] d. p. kingma と j. ba, “adam: a method for stochastic”
0.43
optimization,” CoRR, vol.
とcorr, vol。
0.32
abs/1412.6980, 2015.
abs/1412.6980, 2015
0.35
[52] A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K¨opf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, “Pytorch: An imperative style, high-performance deep learning library,” in NeurIPS, 2019.
A. Paszke氏、S. Gross氏、F. Massa氏、A. Lerer氏、J. Bradbury氏、G. Chanan氏、T. Killeen氏、Z. Lin氏、N. Gimelshein氏、L. Antiga氏、A. Desmaison氏、A.K sopf氏、E. Yang氏、Z. DeVito氏、M. Raison氏、A. Tejani氏、S. Chilamkurthy氏、B. Steiner氏、L. Fang氏、J. Bai氏、S. Chintala氏は2019年、NeurIPSで「Pytorch: 命令型、高性能ディープラーニングライブラリ。 訳抜け防止モード: [52 ]A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. K sopf E. Yang, Z. DeVito, M. Raison, A. Tejani S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, S. Chintala 『Pytorch : 命令型でハイパフォーマンスなディープラーニングライブラリ』 In NeurIPS, 2019。
0.93
[53] C. -C. Chang and C.
53]C。 -C。 Chang と C。
0.50
-J. Lin, “Libsvm: A library for support vector machines,” ACM Trans.
-j。 Lin, “Libsvm: a library for support vector machines”, ACM Trans。
0.55
Intell. Syst.
インテリ。 シスト。
0.54
Technol. , vol. 2, pp. 27:1–27:27, 2011.
テクノル ヴォル。 pp. 27:1-27:27, 2011。
0.50
[54] D. C. Montgomery and G. C. Runger, Applied Statistics and Probability for Engineers.
[54] d.c.モンゴメリーとg.c.ランガーはエンジニアに統計と確率を適用した。
0.63
Applied statistics and probability for engineers, 2014.
2014年、技術者の統計と確率を適用した。
0.58
[55] X. Guo, Y. Li, and H. Ling, “Lime: Low-light image enhancement via illumination map estimation,” IEEE Transactions on Image Processing, vol.
[55] X. Guo, Y. Li, H. Ling, “Lime: 照明マップ推定による低照度画像強調”, IEEE Transactions on Image Processing, vol。