Lightweight Image Deraining Network Based on Dilated Depthwise Separable Convolution and Enhanced Channel Attention

Rainy conditions can introduce significant challenges for various computer vision tasks such as detection and recognition. Addressing the limitations of existing image deraining models in terms of model complexity and performance, this paper proposes a lightweight and efficient deraining network. Firstly, it proposes a multi-scale image feature extraction method based on a shifting window and a fusion method for image features, enhancing the representation of rain streaks. Then, it constructs a separable feature channel attention mechanism to extract rain streak features meticulously, accurately, and efficiently. Finally, it designs a collaborative expansion, depth separable, and globally enhanced detail restoration network to compensate for the loss of some rain streak details that the network may experience. Compared with the traditional image training network, the network significantly reduces the number of parameters, thereby greatly shortening the training time. In addition, the network also showed excellent performance in rainy conditions. Experimental results show that the proposed algorithm performs well on public datasets such as Rain200L, Rain200H and Rain800, and can effectively remove rain artifacts from images.