VDNet: An Image Dehazing Model Based on the Variational Autoencoder
Qihang Sun, Zhihong Yang, Ruifeng Li, Yuyang You · 2024
The hazy weather adversely affects the quality of images from acquisition devices. Traditional image enhancement technology has a poor visual effect. The atmospheric degradation methods combined with prior knowledge are required to restore the image under high-scene conditions. As a solution to these problems, we propose an end-to-end deep learning-based dehazing model VDNet based on the variational autoencoder. We design the flexible feature fusion module to prevent the degradation of low-level features during the model training process. In addition, VDNet introduces the perceptual loss, which effectively extracts the features in line with human perception, making the visual effect of dehazing images better. We use the public hazy image dataset RESIDE for model training and testing. Comparative experiments show that VDNet has excellent visual effects and objective indicators of dehazing with a reduced parameter count of 1.971M, which allows it to be used in devices with limited resources. VDNet achieves the result that the average PSNR is 36.98dB and SSIM is 0.9885.