|
Removing noise from images is a challenging task in front-end computer vision applications, especially in snowy, rainy, foggy, and underwater environments, which pose significant difficulties for various visual tasks. Existing image restoration methods often struggle to achieve robust generalization and are burdened with immense computational complexity. This paper introduces a lightweight image restoration network based on polarized attention mechanisms and efficient feature extraction, capable of addressing tasks like rain removal, snow removal, fog removal, and underwater image enhancement. Initially, a polarized self-attention mechanism is proposed to intricately learn weather noise, followed by an efficient noise removal module complementing the initial network. This addition compensates for any noise information overlooked by the polarized self-attention mechanism. Furthermore, residual connections are integrated into each module of the network, along with grouped convolutions, to prevent network degradation and reduce computational load. In comparison to existing models, this network not only learns noise with similar shapes but also adapts to noise with significant shape variations. Additionally, the introduced residual structure ensures the network's stability, avoiding performance degradation before achieving expected results. The collaborative use of the efficient noise removal module and polarized self-attention mechanism sets a high standard for noise learning in this network. Experimental results demonstrate the efficacy of this algorithm on publicly available datasets including Snow100K, CSD, Rain200H, Rain200L, Rain800, RSID, and EUVP. The proposed method efficiently removes various image noises, effectively achieving the desired image restoration objectives.
|
