In response to the problems of low brightness, weak contrast, noise, and detail loss in low-light images, this paper proposes a low-light image enhancement model, RetinexKIND, which integrates residual dense structure and attention mechanism. It includes three parts: decomposition network, denoising network, and brightness adjustment network. The decomposition network incorporates skip connections and improved Residual Dense Blocks (RDB*) to improve the network’s ability to capture image details and structures, thereby obtaining more accurate illumination and reflection images. The denoising network is constructed by combining the U-Net architecture, Dense Block, and Convolutional Block Attention Module to better enhance the network’s ability to represent image features and increase its focus on specific regions, thereby suppressing noise in reflection images, restoring image details, and improving image quality. The adjustment network consists of convolutional layers and Sigmoid layers with skip connections to improve the contrast of the illumination component. Finally, the denoised reflection component and the enhanced illumination component are fused to obtain a normal light image. Experimental results show that this method improves image brightness, reduces noise, and restores detail information. The PSNR and SSIM are improved by 21.2% and 12.7%.