基于离散注意力机制的信息蒸馏去雨网络研究.

In view of the fact that the images taken in rainy days will be disturbed by noise such as rain stripes, which will affect the processing results of the visual algorithm, an information distillation rain removal network based on discrete attention mechanism is proposed to remove rain from a single image. In this model, the pooled network composed of hole convolution is used for data extraction, and the characteristics of particles of various sizes are integrated to obtain the maximum receptive field and reduce the loss of resolution. The cyclic network is used for selective feature extraction to increase the weight of important features. Through the discrete attention distillation module, the details of the image are gradually refined. After some detail features are obtained, the remaining features are further refined. The algorithm is trained in common data sets rain100 h, rain100 l and self built data set RS1000. The peak signal-to-noise ratio and structure similarityreach 30.06/0.913, 38.61/0.983 and 33.86/0.950 respectively. Experimental data show that the algorithm is superior to the comparison algorithm in both indicators, and the details are preserved more completely. [ABSTRACT FROM AUTHOR]