Résumé: Current shadow removal algorithms generally require prior knowledge. When restoring dark areas or shadow areas with complex textures, there are problems of incomplete restoration of details and distortion. In this paper, a novel image shadow removal algorithm guided by progressive attention mechanism (PAGAN) was proposed for the residual and incomplete shadow removal of complex objects or dark areas. The algorithm combines an attention mechanism with feature fusion technology, which improves the feature location accuracy and enriches the feature information. First, in the feature extraction stage of the generation network, dilated convolution residual blocks were used with different learning rates for feature extraction to expand the receptive field of the network. Then, a parallel attention mechanism was used in the auto-encoder of the generation network to guide the generation network to compile the details of the shadowless image. Second, in the auto-encoding stage, a multi-layer and multi-scale feature fusion method was applied to incorporate the global semantic information and local detail features. Finally, the series attention mechanism was used to guide the discrimination network to identify the shadowless image generated by the generation network to reduce the loss of key features and enhance the identification ability of the discrimination network. Experiments based on open datasets, SRD and ISTD, indicated good visual effect. The structural similarity index measure value of the algorithm can reach 97.5 %, the peak signal-to-noise ratio can reach 32.8, and the root mean square error can be reduced to 5.8.

Langue: Anglais