SEGAN: Unsupervised Image Enhancement Based on Edge Information Guidance

To address issues such as the lack of paired images and insufficient preservation of image details in low-light enhancement, this paper investigates a low-light image enhancement algorithm based on an unsupervised Generative Adversarial Network (GAN) guided by edge information. This algorithm enables unsupervised training even in the absence of low-light/normal-light image pairs. In the generator component, a Self-Regularized Attention Map (SAM) incorporating mixed edge information and the HaLoAttention local self-attention mechanism are introduced into the U-Net architecture. The Structure-guided Self-feature Guided Enhancement (SSGE) module incorporates an edge information extraction module (Edge Extraction, EX) that fuses attention weights and an edge information enhancement module (Edge Information Enhancement, EIE) based on dynamic convolution and Spatially-Adaptive Normalization (SPADE). The discriminator component adopts a global-local discriminator structure, randomly cropping from the generator's enhanced results and unpaired normal light images, both of which are subjected to true/false identification using PatchGAN. Experimental results show that on the reference test set composed of LOLv2 and LSRW, the proposed algorithm outperforms competing methods in PSNR, SSIM, and NIQE (with improvements of 0.8 dB, 0.02, and 0.3 over the best baseline). In addition, on the unreferenced datasets DICM, our method also achieves the best performance in NIQE, producing enhancement results that are closer to real-world effects.