EMNet: Efficient Multi-Scale Convolutional Network for Image Restoration
Yaoze Lian, Liang Zhou · 2024
With the rise of deep learning, numerous methods based on convolutional neural networks have emerged in various fields of image restoration. Neural networks and recent advancements like transformers have performed exceptionally well in many visual tasks. In this paper, we analyze the strengths and weaknesses of convolutional methods and transformer approaches. We integrate a Multi-scale Enhanced Convolution module with dilation mechanisms and a Multi-shape Feature Enhanced Unit with various attention combinations into the MixBlock, effectively achieving both a large receptive field and local information interaction. Additionally, our model leverages a U-shaped architecture, enabling it to effectively capture multiscale representations. We introduce the network as EMNet-a highly efficient CNN architecture specifically designed for image restoration while ensuring computational efficiency. Extensive experiments demonstrate that our network achieves state-of-the-art performance on three benchmark datasets, including image dehazing, desnowing, and motion deblurring.