Single image de-raining via clique recursive feedback mechanism

Abstract Deep learning methods have achieved significant progress in single image rain removal in recent years. However, the remaining rain streak phenomenon still exists in previous ways, provided that they did not adequately consider the interaction of features flowing among the network. To utilize the feature interaction of the network, we propose a new rain removal network based on the clique recursive feedback mechanism. Mainly, considering the interaction of the feature between different convolution layers, we construct a residual clique block (RCB) to infer the local information, allowing the network to rectify the model parameters in RCB alternately. Besides, a multi-path dilated convolutional unit is embedded into a scale clique block (SCB) to cover more scale components. In SCB, we consider the complementary correlation of different scales, and the multi-scale features are updated alternately, which is essential for excellent feature representation. Along with the clique recursive feedback mechanism, the information flowing among RCB and SCB is thus maximized during propagation. Experiments on synthetic and real-world images have shown the superiority of the proposed method over state-of-the-art methods.