Abstract: Aiming at the existing blind image quality evaluation algorithms with low performance in the face of real distorted images, this paper proposes the multi-level feature fusion and semantic information enhancement for NR (MFFSE-NR), which combines multi-layer feature fusion and semantic enhancement for NR. The local and global distortion features of the image are extracted, and the features are fused in multiple layers by using the feature fusion module; the semantic information is enhanced by using the multilayer dilation convolution, which then guides the mapping process from distorted images to quality scores; considering the relative ranking relationship between predicted scores and subjective scores, the L1 loss function and the ternary ranking loss function are fused to construct the new loss function Lmix. in order to validate the method, validation and comparison experiments are carried out on the LIVEC dataset, and the SROCC and PLCC metrics of the algorithm are improved by 2.3% and 2.3%, respectively, compared with the original algorithm; cross-dataset experiments are carried out on the KonIQ-10k dataset and the LIVEC dataset, and the algorithm exhibits a good generalization performance in the face of real distorted images.
Download 2023.7.31