针对传统Retinex算法存在的增强图像纹理不清晰、适用光线强度单一等问题，提出了一种基于残差和多尺度注意力机制的Retinex低光图像增强算法。首先使用引入残差模块和跳跃连接的U-Net结构确保模型能完整地提取特征信息，以及将原始图像分解为准确的光照分量和反射分量；然后采用结合多尺度注意力机制的恢复网络对反射分量进行处理，提高网络对退化信息的感知能力和对细节纹理信息的抓取能力；接着使用调整网络增强光照分量的光线强度；最后融合处理后的反射分量与照度分量，得到最终的输出图像。将所提算法与7种同类型算法进行比较，主观结果和客观评价均表明经过所提算法的增强结果在对比度增强、噪声处理、色彩自然度方面取得了优异的表现。实验结果表明，本文中所提方法能有效提高图像对比度、抑制噪声，具有符合人眼观感的视觉表现，且能有效作用于各种光照环境，为下一步的图像处理提供了可靠的信息源。

A Retinex-based low-light image enhancement algorithm based on residual and multi-scale attention mechanisms was proposed to address the issues of unclear texture enhancement and the limitation of single light intensity in traditional Retinex algorithms. Initially, a U-Net structure with residual modules and skip connections was utilized to ensure that the model can fully extract feature information and accurately decompose the original image into illumination and reflectance components. Subsequently, the recovery network combined with the multi-scale attention module was applied to process the reflectance component, enhancing the network's perception of degraded information and its ability to capture texture detailed information. Then, the illumination component's light intensity was enhanced by the adjustment network. Finally, the processed reflectance component and illumination component were fused to obtain the final output. Comparative subjective and objective evaluations with seven similar algorithms indicated that the enhancement results achieved by the proposed algorithm exhibit excellent performance in contrast enhancement, noise reduction, and natural color rendering. The experimental results demonstrate that the method presented in this paper effectively improves image contrast, suppresses noise, and provides a visually appealing presentation that aligns with human visual perception. Moreover, it is effective in various lighting environments, offering a reliable source of information for subsequent image processing.