A Lightweight Image Dehazing Algorithm Based on Deep Learning

The dehazing algorithm based on deep learning has good performance. However, this way has some problems such as high computational complexity and long latency time still exist, making it difficult to apply in real-time systems. To address this issue, this paper proposes a lightweight network called zUNet based on UNet. Firstly, a lightweight convolution module is designed to obtain redundant feature maps through low-cost computation. Then, the channel attention mechanism is used to extract global features. Replace The convolution modules in UNet are replaced with gate mechanism residual blocks to dynamically fuse feature maps. Finally, the proposed zUNet algorithm is compared with mainstream algorithms such as FFA-Net on the SOTS-IN and SOTS-OUT datasets. The results show that the peak signal-to-noise ratio (PSNR) of zUNet reaches an average of 37.38 dB, which is the best among all the evaluated algorithms. The number of parameters (Param) are 0.552M, the multiply– accumulate operations (MACs) are 1.933G, and the latency time is 6.031ms, at least 42%, 90%, and 39% lower, respectively.