Abstract
Dehazing can improve the clarity of images and provide more reliable inputs for image analysis tasks, thereby enhancing their performance. Therefore, we propose a dehazing network based on knowledge transfer and multi-data enhancement correction. First, we propose a multi-data enhancement correction method that combines different image enhancement techniques to improve the quality of the input images. Second, by leveraging a pre-trained teacher network to acquire prior knowledge from clear data, guiding the dehazing process of the student network through knowledge transfer. We introduce a deep multi-scale refinement network composed of a dense feature enhancement module and enhanced residual dense blocks, enabling the dehazing model to learn the local structure and feature representation of the data more accurately. Experimental results on multiple benchmark datasets demonstrate that the proposed dehazing method outperforms state-of-the-art dehazing methods. Code are available at: https://github.com/JNcmm/KTMDA-DehazeNet.