CPDINet: Blind image quality assessment via a content perception and distortion inference network

Abstract

Nowadays, it is still challenging for the blind image quality assessment (BIQA) to accurately predict quality scores of distorted images, since distorted images have rich content information and complex distortions. To solve these problems, a content perception and distortion inference network for BIQA is proposed, which divides IQA task into content perception and distortion inference processes. Since humans try to understand image content before perceiving quality scores, a content feature extractor is designed to explore content information in an image to deal with the content variation problem. To handle the distortion diversity problem, a distortion feature extractor is proposed to capture distortion features in images. Because extracted content features and distortion ones have different characteristics, attention-based fusion blocks to fuse multi-scale content features and distortion ones as guidance to selectively enhance important features based on calculated weight scores are proposed. With fused features, a quality prediction module is designed to regress multi-scale features to quality scores. Experiments are performed on six public IQA datasets, including LIVE, CSIQ, TID2013, LIVEC, KonIQ-10k, and SPAQ. Experimental results show that the proposed method can effectively predict quality scores for both synthetically and authentically distorted images than its peers, including the state-of-the-art methods.

No‐reference image quality assessment via a dual‐branch residual network

“…In this section, to validate the effectiveness of the proposed method, we compare it with several state‐of‐the‐art FR‐IQA methods (PSNR [42], SSIM [43], and FSIMc [44]) and NR‐IQA methods (BRISQUE [7], CORNIA [10], M3 [24], HOSA [25], FRIQUEE [26], IQA‐CNN [15], Prewit [27], CaHDC [18], HyperIQA [28], DBCNN [29], MMMNet [30], VCRNet [31], and CPDINet [32]) on four public databases. The SROCC and PLCC results are shown in Table 2, with bold highlighting for top‐ranking results.…”

“…The results of three FR‐IQA methods (PSNR [42], SSIM [43], and FSIMc [44]), four handcrafted feature‐based NR‐IQA methods (BRISQUE [7], CORNIA [10], HOSA [25], and FRIQUEE [26]), and five deep learning‐based NR‐IQA methods (IQA‐CNN [15], CaHDC [18], HyperIQA [28], DBCNN [29], and VCRNet [31]) were calculated using the source code provided by the respective authors. Due to the absence of code for M3 [24 ]、Prewit [27 ]、MMMNet [30], and CPDINet [32], all results are derived from the existing literature.…”

“…This approach utilizes a non-adversarial model to effectively handle the quality assessment task for distorted images. Shao et al divided the IQA task into content perception and distortion inference processes [32]. They designed content feature extractors and distortion feature extractors to capture content information and distortion features separately from images.…”

“…Most existing NR-IQA methods [13,18,32] typically take individual image blocks cropped from distorted images as separate inputs to the model and then average their quality scores to estimate the overall image quality score. These methods overlook the overall nature of a distorted image, meaning that the quality of an image block depends not only on its own features but also on the influence of other image blocks within the same image.…”

“…and CPDINet [32]) on four public databases. The SROCC and PLCC results are shown in Table 2, with bold highlighting for top-ranking results.…”