Fscope-IQA: A Novel Attention Design for Image Quality Assessment

The existing algorithms for no-reference image quality assessment (NR-IQA) always extract deeper features from low-level features, which often tend to the attention on salient objects. However, the purpose of IQA is not to evaluate the object-related regions merely but the entire image scope. Therefore, we design a full-scope attention guided image quality assessment algorithm (Fscope-IQA), in which there are three key elements: backbone, non-object attention features extraction module (NAFE) and cross-guided features coupling module (CGFC). We adopt the vision Transformer as our backbone. In the NAFE, the non-object attention features are acquired by the difference of the deep and shallow features. In the CGFC, a spatial mutual fusion is produced by exchanging and coupling distorted information between the object and non-object features. The full-scope information is obtained by repetitive combination of NAFE and CGFC, to compute the quality score in line with human vision. We performed experiments on both authentic and synthetic distortion datasets, and results show that our method could reach satisfactory performance over the up-to-date methods. The analysis illustrates that the full-scope attention is much more reasonable for NR-IQA than the object-oriented attention.