Non-Uniform Dehazing Algorithm Based on Improved ConvNeXt

Haze generally leads to a degradation of image quality, characterized by reduced contrast, color cast, and structural distortion. We observe that many deep learning-based models exhibit remarkable performance in removing uniform haze, but they typically fall short in handling non-uniform haze images. There are two main factors contributing to this issue. The first factor is that the degree of haze coverage varies across different regions of the image, making it challenging to maintain a consistent color balance in the restored image, with the imbalance directly proportional to the unevenness of the haze. The second factor is that non-uniform haze images, limited in data scale, are difficult to train effectively using convolutional neural networks (CNN). Addressing these issues, we propose an enhanced ConvNeXt model that is based on two-dimensional Discrete Wavelet Transform (DWT), FFC, and the pre-trained original ConvNeXt model, with a parallel implementation of Transformers to further deepen the network. This approach improves the dehazing capability and has achieved higher Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) across multiple datasets.

1. Introduction

Haze is a natural phenomenon of nature that affects visibility, causing observed objects to appear blurred and obscure. Under such conditions, many intelligent algorithms or applications are impacted and hindered, including but not limited to target detection[l], autonomous driving[2], etc. This underscores the importance of image dehazing in scientific research areas such as computer vision[3]–[7] [20] as well as in production and daily life.