An improved CycleGAN model for low-light or exposure-light images preprocessing and its application

Image preprocessing model plays an improvement role in accurately detecting objects in low-light or exposure-light images, because it can provide clear images for image detection models. An improved CycleGAN (ICGAN) model is presented to enhance low-light or exposure-light images in this paper. Under the framework of CycleGAN, a bidirectional attention mechanism is added to focus on object features area on images, and the convolutional structures are designed to enhance local perception ability before down sampling the feature maps. The ICGAN model is applied to preprocess the low-light or exposure-light surfaces images of hot-rolled steel strips. The experiment results show that compared with CGAN, the average gradient, the information entropy, and the PSNR index of ICGAN have been increased, but the SSIM index of ICGAN has been slightly reduced.