HCL-Net: A Single Image Rain Removal Algorithm Based on Channel Attention and Multi-scale Feature Fusion

The investigation of rain removal from pictures has always been a pivotal matter in the domain of computer vision. Nevertheless, the majority of existing algorithms employ encoder-decoder structures that have been trained on artificially synthesized datasets, without taking into account the harm that heavy rain can cause to actual images. Batch normalization is unsuitable for low-level computer vision duties, and skip connections in encoder-decoder architectures are not always appropriate for rain removal. This has resulted in setbacks in relevant domains. In this article, we propose a groundbreaking single-image rain removal algorithm that is founded on channel attention and multi-scale feature fusion. Initially, we build a contextualized dilated restoration block to restore the rain-affected image during the feature extraction phase. Secondly, we create a half-instance normalization residual block to conserve more scale information and to augment the modeling capacity of the model, without sacrificing performance while hastening model convergence. Finally, in response to the view that simple skip connections may not always be advantageous, we have constructed a LeakyReLU channel attention block to aggregate information and enhance the effectiveness of skip connections. Our comprehensive experiments on authentic rain datasets and images provide evidence that our model can more proficiently remove rain streaks and accumulations from pictures, in contrast to the existing methods.