Biologically inspired image quality assessment
Introduction
The past decades have witnessed a dramatic increase in the number of images with the tremendous development of social networking websites, smartphones, and cameras. And various systems have been developed to deal with such a large scale of images. In these systems, image quality usually plays a significant role. For example, images of poor quality may lead to obstacles in learning or applying such systems for practical applications, e.g. scene recognition [1], image retrieval [2], and so on. In addition, image quality can be adopted as a criterion for evaluating the performance of image processing systems [3], [4], [5], optimizing image processing algorithms, and monitoring the working condition of devices [6]. Thus it is meaningful to develop image quality assessment (IQA) methods that can precisely and automatically estimate human perceived image quality.
In recent years, many IQA methods have been developed and we can classify them into three classes [6]: full-reference (FR) IQA [7], reduced-reference (RR) IQA [8], [9], and no-reference (NR) or blind IQA [10], [11]. FR-IQA methods need all the information of the reference image, i.e. the undistorted version of the test image, is needed. In contrast, RR-IQA and NR-IQA methods only need part of or none of the information about the reference image. Consequently, the quality prediction accuracies of FR-IQA methods are usually better than present RR-IQA and NR-IQA methods.
Generally speaking, the intrinsic idea of FR-IQA is to estimate the quality of a test image by measuring the similarity or difference between the test image and the corresponding reference image. For example, in peak signal-to-noise ratio (PSNR) and root mean squared error (RMSE), the most widely used two IQA methods, the differences between the reference image and the test image are calculated pixel by pixel, and then combined into a single value. Because PSNR and MSE do not always consistent well with human perception [12], great efforts have been paid to develop progressive methods for quality assessment in the past decades [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28]. And many of them have shown impressive and inspiring consistency with human perception over a large range of datasets [26], [27], [28].
Since the goal of IQA is to approximate human beings’ judgments of image quality, it is meaningful to develop IQA methods that mimic the perception mechanism of the human visual system (HVS). Although many attempts have been proposed, most of them only consider some particular properties of HVS, e.g. contrast sensitivity function (CSF) [18], [19], [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], just noticeable difference (JND) [30], and visual attention (VA) [31], etc. Usually they do not perform as well as state-of-the-art FR-IQA methods. To date, only limited IQA methods have been proposed to formula the processing in the visual cortex [27]. And the properties of primary visual cortex, V1, have not been well explored for IQA, although neuroscientists have demonstrated that V1 plays a much significant in visual processing [32].
In this paper, we utilize biologically inspired feature (BIF) models [33] to mimic the properties of (S1) and complex (C1) cells in V1, and construct a novel IQA index by measuring the similarity between the BIFs of the reference image and those of the test image. Although BIFs have been introduced to FR-IQA before, it was adopted for estimating visual attention [34]. In contrast, in the proposed method, BIFs are deplored for representing the input image in the primary visual cortex and directly utilized for quality prediction. Thorough experiments conducted on various IQA databases demonstrate that the proposed method is in highly consistency with human perception and outperform state-of-the-art FR-IQA methods across a number of datasets. The highlights of the proposed method are summarized below as follows:
- a.
We explore BIF for FR-IQA by employing it to mimic the processing in the primary visual cortex;
- b.
We construct a novel FR-IQA framework by measuring the similarity between the BIFs of the test image and the BIFs of the corresponding reference image; and
- c.
Thorough experiments on existing databases demonstrate that the proposed method is highly comparable with state-of-the-art FR-IQA methods.
- d.
The rest of the paper is organized as follows. Section II introduces the calculation of BIFs. In Section III, we present the framework of the proposed quality evaluation method. Extensive experiments conducted on standard IQA datasets are presented and analyzed in Section IV. Section V concludes the paper.
Section snippets
Biologically inspired features
Biologically inspired feature models mimic the tuning properties of the simple and complex cells in V1 and have been demonstrated to be effective and efficient for solving various image processing problems, e.g. scene classification [33], object recognition [35], [36], visual attention detection [32], [33], [34], [35], [36], [37], and so on. We thus choose to use BIF for representing an image in the proposed research. Specially, we follow the work presented in [33], and adopt the C1 units,
Quality assessment framework
The proposed IQA approach simulates the process in the visual cortex and can be divided into three main components: the biologically inspired feature maps, the similarity maps between the distorted feature maps and the relevant reference feature maps, and the percentile pooling based quality prediction. And we term the proposed IQA method as biologically inspired feature similarity (BIFS). The flowchart of the BIFS is shown in Fig. 2. Details will be discussed below.
Experimental results
To evaluate the performance of the proposed method, we test it on several existing IQA databases and compare it with a number of state-of-the-art FR-IQA methods.
Conclusions
In this paper, biologically inspired feature is introduced to mimic the processing in the primary visual cortex. Afterwards, the similarity between the BIFs of the reference image and the BIFS of the distorted image is calculated for quality prediction. The comparison of both proposed algorithm with state-of-the-art FR-IQA metrics on a number of databases shows it has an impressive consistency with human perception and overwhelming superiority over the state-of-the-art FR-IQA metrics for
Acknowledgements
This paper is supported by the National Natural Science Foundation of China (No 61472110), the Program for New Century Excellent Talents in University (NECT-12-0323), the HongKong Scholar Programme (XJ2013038), Zhejiang Provincial Natural Science Foundation of China (No. LR15F020002).
References (49)
- et al.
Semantic embedding for indoor scene recognition by weighted hypergraph learning
Signal Process.
(2015) - et al.
Color-to-Gray based on chance of happening preservation
Neurocomputing
(2013) - et al.
Multi-view intact space learning
IEEE Transactions on Pattern Analysis and Machine Intelligence
(2015) - et al.
Learning to rank using user clicks and visual features for image retrieval
IEEE Trans. Cybern.
(2015) - et al.
Video tonal stabilization color states smoothing
IEEE Trans. Image Process.
(2014) - et al.
Color to gray: visual cue preservation
IEEE Trans. Pattern Anal. Mach. Intell.
(2010) - et al.
Modern Image Quality Assessment
(2006) - et al.
A statistical evaluation of recent full reference image quality assessment algorithms
IEEE Trans. Image Process.
(2006) - et al.
Quality-aware images
IEEE Trans. Image Process.
(2006) - et al.
Image quality assessment based on multiscale geometric analysis
IEEE Trans. Image Process.
(2009)
Sparse representation for blind image quality assessment
Proc. IEEE Conf. Comput. Vis. Pattern Recognit. (CVPR)
Universal blind image quality assessment metrics via natural scene statistics and multiple kernel learning
IEEE Trans. Neural Netw. Learn. Syst.
Mean squared error: love it or leave it?-A new look at signal fidelity measures
IEEE Signal Process. Mag.
Visibility of wavelet quantization noise
IEEE Trans. Image Process.
A wavelet difference predictor
IEEE Trans. Image Process.
Image quality assessment: from error visibility to structural similarity
IEEE Trans. Image Process.
An information fidelity criterion for image quality assessment using natural scene statistics
IEEE Trans. Image Process.
Image information and visual quality
IEEE Trans. Image Process.
VSNR: a Wavelet-based visual signal-to-noise ratio for natural images
IEEE Trans. Image Process.
A content-based image quality metric
Information content weighting for perceptual image quality assessment
IEEE Trans. Image Process.
A novel metric based on MCA for image quality
Int. J. Wavel. Multiresolution Inf. Process.
Image quality assessment based on S-CIELAB model
Signal Image Video Process.
Color Fractal Structure Model for Reduced-reference Colorful Image Quality Assessment
Cited by (62)
Representation learning of image composition for aesthetic prediction
2020, Computer Vision and Image UnderstandingCitation Excerpt :To this end, exploring hierarchical deep features (Jun et al., 2019; Yu et al., 0000; Gao et al., 2018) and metric learning (Jun et al., 2017a) are promising solution. In addition, photo aesthetic is highly correlated with photo fidelity, i.e. traditional image quality (Gao and Yu, 2016; Gao et al., 2017). In the future, we will pay efforts to combine both aesthetic and fidelity indices for image restoration and enhancement.
Attentive and ensemble 3D dual path networks for pulmonary nodules classification
2020, NeurocomputingCitation Excerpt :Traditionally, researchers explore hand-crafted features and use a classifier to predict the category of a nodule [2–4]. Nowadays, deep learning techniques, especially Convolutional Neural Networks (CNNs) have achieved great success in various high-level visual understanding tasks [5–10]. Researchers are thus inspired to employ CNNs for pulmonary nodule classification.
LPR-Net: Recognizing Chinese license plate in complex environments
2020, Pattern Recognition LettersMulti-task learning for object keypoints detection and classification
2020, Pattern Recognition LettersIntegrating prediction and reconstruction for anomaly detection
2020, Pattern Recognition Letters