Aiming at the low performance of the existing blind image quality assessment algorithm when facing the real distorted images,the paper proposes a new no-reference image quality assessment algorithm,namely multi-level feature fusion and semantic enhancement for NR(MFFSE-NR),which combines multi-level feature fusion and semantic in-formation enhancement.The local and global distortion features of an image are extracted,then a feature fusion module is used to fuse the features in layers.The multi-layer dilated convolution is employed to enhance semantic information and further direct the mapping process from distorted image to quality fraction.Finally,a novel loss function called Lmix is created by combining the triplet ranking loss function and the L1 loss function,taking account of the relative ranking relationship between the predicted score and the subjective score.Validation and comparison experiments carried out on LIVEC dataset show that both the SROCC and PLCC index are improved respectively by 2.3%than the original al-gorithm;cross-dataset validation on the KonIQ-10k dataset and LIVEC dataset confirm that the proposed algorithm has good generalization ability when dealing with the real distorted images.
deep learning/image quality/convolution neural network/feature extraction/channel attention structure/multi-level feature fusion/dilated convolution/triplet loss function
维普
万方数据