The paper presents a two-stage depth-guided perception framework, UVZ, for underwater image enhancement (UIE).
In the first stage, UVZ includes a depth estimation network (DEN) and an auxiliary supervision network (ASN). DEN generates a depth map that captures the degradation characteristics of the underwater scene, while ASN promotes the correlation between the depth map and the input image during training.
In the second stage, UVZ employs a depth-guided enhancement network (DGEN) that consists of two branches: a swin transformer-based non-local branch and a multi-kernel convolution-based local branch. The non-local branch captures long-range dependencies, while the local branch focuses on nearby features. The depth map guides the selective fusion of these two branches, enabling adaptive enhancement of near and far regions.
The proposed UVZ framework outperforms state-of-the-art UIE methods in both qualitative and quantitative evaluations on multiple benchmark datasets. Additionally, UVZ exhibits good generalization abilities in various visual tasks, such as image segmentation, keypoint detection, and saliency detection, without any parameter tuning.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문