SGIQA: Semantic-Guided No-Reference Image Quality Assessment

Abstract

Existing no reference image quality assessment(NR-IQA) methods have not incorporated image semantics explicitly in the assessment process, thus overlooking the significant correlation between image content and its quality. To address this gap, we leverages image semantics as guiding information for quality assessment, integrating it explicitly into the NR-IQA process through a Semantic-Guided NR-IQA model(SGIQA), which is based on the Swin Transformer. Specifically, we introduce a Semantic Attention Module and a Perceptual Rule Learning Module. The Semantic Attention Module refines the features extracted by the deep network according to the image content, allowing the network to dynamically extract quality perceptual features according to the semantic context of the image. The Perceptual Rule Learning Module generates parameters for the image quality regression module tailored to the image content, facilitating a dynamic assessment of image quality based on its semantic information. Employing the Swin Transformer and integrating these two modules, we have developed the final semantic-guided NR-IQA model. Extensive experiments on five widely-used IQA datasets demonstrate that our method not only exhibits excellent generalization capabilities but also achieves state-of-the-art performance.