Contextual Information Aggregation and Multiscale Feature Fusion for Single Image De-Raining in Generative Adversarial Networks

Abstract

Abstract Needle image de-rain the existence of non-uniform rain density, noise misjudgment, and other problems, proposed a context information aggregation and multi-scale feature fusion generative adversarial network(CMGAN) single image de-rain method. Firstly, design the encoding, context information aggregation, and decoding structures to form the generator network, extract features by convolution, expand the convolution to aggregate context information effectively, and transposing convolution to restore the image. Enhance the model's capability to perceive image details, enabling it to accurately analyze image information and reconstruct image content.; secondly, design the multiscale feature fusion discriminator network, and capture the image's different details through the convolution kernel of different scales. Enhance the model's capacity to capture image details by integrating feature maps from various scales, enabling it to effectively discern between authentic and manipulated images.; finally, a new refinement loss function is proposed to reduce the grid artifact generation, and Lipschitz constraints are added to further reduce the imaging gap. In this study, the model's performance is assessed using peak signal-to-noise ratio and structural similarity as evaluation metrics, and the experiments undertaken on real and synthesized rain images show that the method exhibits excellent rain removal performance.