Low-light Image Enhancement Based on Conditional Generation Adversarial Network

Images taken in low-light environment generally suffer from issues such as high noise level, low contrast, and color distortion. These problems not only affect the visual quality of the image, but also hinder the completion of subsequent machine vision tasks. In this paper, we present a low-light image enhancement method based on conditional generation adversarial network. The network’s generator utilizes residual dense blocks coupled with channel attention mechanisms to boost feature extraction’s precision and swiftness. Residual dense blocks are employed to fortify feature propagation and reuse, whereas channel attention mechanisms selectively recalibrate channel weights, thereby enhancing the efficiency of image processing. Besides, we employ a PatchGAN architecture as the discriminator to assess image quality. For training optimization, we introduce a multi-module loss function that considers local texture, global content, structural clarity, content fidelity and color authenticity. The qualitative and quantitative experiments demonstrate that our method balances processing time and effect, achieving tasks such as noise suppression, contrast enhancement, and color restoration.