In order to improve the representation of network on features,a novel optical remote sensing salient target detection algorithm based on feature refocusing and refinement is proposed. Adjacent layer features are used to capture the context semantic complementary information interactively,and the range of information extraction from receptive field is adjusted by expanding convolution to complete the initial feature focusing. Then,the attention mechanism is applied to the deep layer features to form a location guidance module to enhance the attention to the salient features and complete feature refocusing. Finally,the attention map and anti attention map of salient features are obtained through shallow layer features,the network is guided to further mine the information of high confidence salient regions and low confidence background regions,and the optimized features are refined. Two open datasets,EORSSD and ORSSD,are used for experiment and evaluation to prove the effectiveness of the algorithm.
维普
万方数据