Visual foundation models (VFMs) have demonstrated impressive generalization capabilities in computer vision tasks. Previous studies show that fine-tuning VFMs with learnable tokens can achieve better generalization performance than full-parameter fine-tuning. The problem we need to address is how to learn the tokens that focus on the content information while ignoring the influence of style. For this purpose, we propose a novel Dual-Branch Content-enhanced Token (DBCT) learning framework. Specifically, we construct a style-suppressing branch, which contains a Style-sensitive Channel Suppression (SCS) module to transform the frozen VFM features into style-suppressed features, enabling the learning of style-invariant tokens. In addition, to compensate for the content degradation caused by the style-suppressing branch, we introduce a content-preserving branch that directly takes the frozen VFM features as input to learn content-focused tokens. Meanwhile, we propose a Token-query Linking (TLink) strategy to connect the two sets of tokens with the queries in the decoder. Through extensive experiments, our method achieves advanced results on various benchmarks.

Captured retinal images usually have quality issues, manifested as containing multiple distortions (e.g., low light and blurring). Low-quality images bring a challenge to the screening and diagnosis of ophthalmic diseases. Existing image enhancement methods typically neglect the analysis of distortions and require high-quality reference images for model learning, making them unsuitable for clinical applications. In this paper, we propose a Distortion-Aware Network (DANet) for retinal image enhancement in a zero-reference way. DANet consists of three parallel branches by incorporating atmospheric scattering theory, which decomposes the low-quality image into a clean image, a transmission map, and an atmospheric light map. The upper branch utilizes a dark channel prior module to estimate the atmospheric light map, and the middle branch uses a transmission map generation module to estimate the transmission map. In contrast, the lower branch uses a deblurring module and a low-light enhancement module to obtain a deblurred image and an illumination-enhanced image and fuses these two images using a fusion block to generate the final enhanced image. Taking into account the limited publicly available datasets, we curate two datasets for the retinal image enhancement task. Experimental results show that our DANet can greatly improve the visual quality of the image with good interpretability, achieving superior performance over seven state-of-the-art methods.