LEESDFormer: A lightweight unsupervised CNN-Transformer-based curve estimation network for low-light image...

Current low-light image enhancement methods mainly focus on improving the low-light regions within images. However, they often fail to adequately consider the impact of mixed exposures and noise on the images, resulting in suboptimal enhancement results and even loss of some detailed information. Moreover, these methods predominantly rely on convolutional neural networks (CNNs), which have inherent constraints in capturing long-range dependencies and global information. To address these issues, this paper introduces LEESDFormer, the first unsupervised low-light image enhancement method based on CNN-Transformer. Firstly, we propose a Low-light Enhancement and Exposure Suppression S-shaped curve (LEES-S curve), which simplifies the complex challenge of low-light enhancement and exposure suppression into a simpler curve estimation task, thus substantially reducing the task's complexity. LEESDFormer iterates the LEES-S curve through the Low-light Image Enhancement and Exposure Suppression Module (LEESM), thereby achieving desired enhancement effects. Subsequently, the Image Denoising Module (IDM) is employed to denoise the enhanced images. Extensive experiments demonstrate that our method exhibits excellent robustness, generalization capabilities, and visual effects compared to state-of-the-art unsupervised low-light image enhancement methods, even outperforming some supervised learning approaches. Notably, our method achieves a Peak Signal-to-Noise Ratio (PSNR) of 21 dB on the LOL-v2-real dataset, demonstrating its superior enhancement performance and denoising capability. Furthermore, LEESDFormer is simple and efficient, with only 65 K parameters, and processes each image in merely 8 ms, making it deployable on resource-limited devices and having significant practical value.