Effective Image Dehazing via Wavelet-Transform and Attention Mechanism
Haotian Jiang · 2024
Dehazing represents a critical image processing technique primarily aimed at addressing the issues of low contrast and blurred details in images resulting from fog. Despite advancements in technology leading to the development of numerous algorithms, many of these still fall short in fulfilling the efficiency demands of practical applications. The integration of wavelet technology has provided novel perspectives on dehazing techniques; in contrast to traditional dehazing algorithms, wavelet-based methods demonstrate a greater capacity to mitigate over-smoothing and artifacts, thereby exhibiting significant potential in image classification, recognition, and traffic monitoring. This paper presents a dehazing algorithm that integrates wavelet transforms and the attention mechanism, aimed at addressing the efficiency deficiencies associated with image dehazing. This paper employs a framework similar to that of U-net and utilizes the reversible properties of the discrete wavelet transform to facilitate the transfer of intermediate features. Additionally, a novel approach is introduced in the downsampling module, optimizing the feature downsampling process through the implementation of an attention mechanism. This mechanism improves the ability to manage long-range dependencies and adaptively aggregate spatial information, leading to increased convergence speed and inference efficiency. Concurrently, it enhances overall performance and ensures high-quality image reconstruction. Moreover, the model incorporates a refinement module based on Retinex theory, which serves to restore the color and detail of the images. Experimental results across several extensive dehazing datasets indicate that our method not only surpasses many state-of-the-art techniques in single image dehazing performance but also significantly reduces computational costs.