Fast image dehazing algorithm based on multi-scale module with multi-level receptive field
Haobin Liu, Linhui Fu · 2025
Deep learning has shown promising results in image dehazing tasks, but the large number of parameters typically involved affects processing speed. To balance dehazing quality and speed, this paper proposes a multi-level lightweight scale aggregation network. By leveraging an adaptive mechanism that combines pooling and convolutional layers, the model expands the receptive field efficiently while minimizing parameter overhead. Furthermore, a pixel dehazing module is designed to map hazy image pixels to clear ones through pixel-wise feature processing. Tested on the indoor RESIDE dataset, the proposed algorithm achieves significantly faster dehazing speeds compared to benchmark models. It also ranks first in quantitative metrics, with PSNR and SSIM values of 23.62dB and 0.9310, respectively.