| Low-light image enhancement is a complex task in computer vision,as it involves addressing multiple types of degradation beyond just improving brightness,such as color distortion,noise,detail loss,shadow blocks,and halo artifacts.To tackle these chal-lenges,this paper proposes three deep learning networks inspired by visual perception and attention mechanisms.Firstly,this paper introduces the Lightweight Stacked Atten-tion Residual Network for Low-Light Image Enhancement.This network stands out due to its small model parameters,lightweight architecture,fast processing speed,and promising enhancement performance,making it suitable for real-world scenarios.Secondly,this pa-per presents the Two-Stage Network with Channel Attention,which effectively enhances brightness in low-light images and restores the enhanced images while addressing var-ious types of degradation.By leveraging both RGB and HSV color spaces,image en-hancement is performed in the HSV space followed by restoration in the RGB space.Thirdly,this paper proposes the Degradation-Aware Deep Retinex Network based on an improved Retinex Theory.This network optimizes traditional Retinex algorithms and accelerates Retinex-based enhancement methods,thereby achieving efficient low-light image enhancement.this paper extensively trains and evaluate our proposed deep learn-ing models using LOL datasets,which include both real-world and synthetic data.Addi-tionally,this paper tests the performance of our models on challenging low-light image enhancement datasets(LIME,DICM,MEF,NPE)without Ground-Truth.Through ex-tensive ablation and comparative experiments,our models demonstrate promising results(PSNR>22d B,SSIM>0.82)in various quantitative and qualitative metrics.Our proposed models surpass many state-of-the-art methods,achieving highly promising performance.Our model simulates the attention mechanism and perception mechanism of human vision and achieves good objective index effect and visual perception effect. |
PDF Full Text Request