Abstract
Image dehazing models are critical in improving the recognition and classification capabilities of image-related artificial intelligence systems. However, existing methods often ignore the limitations of receptive field size during feature extraction and the loss of important information during network sampling, resulting in incomplete or structurally flawed dehazing outcomes. To address these challenges, we propose a multi-level perception fusion dehazing network (MPFDN) that effectively integrates feature information across different scales, expands the perceptual field of the network, and fully extracts the spatial background information of the image. Moreover, we employ an error feedback mechanism and a feature compensator to address the loss of features during the image dehazing process. Finally, we subtract the original hazy image from the generated residual image to obtain a high-quality dehazed image. Based on extensive experimentation, our proposed method has demonstrated outstanding performance not only on synthesizing dehazing datasets, but also on non-homogeneous haze datasets.