关键词: 深度学习, 卷积神经网络, 计算机视觉, 图像增强

Abstract: Tunnel images, affected by the shooting environment, suffer from uneven illumination distribution, local occlusion, and many noises. To address the problems of overexposure and distortion in existing image enhancement algorithms, this paper proposes a tunnel image enhancement algorithm called DA-Zero-DCE (Denoising-Attention based Zero-Reference Deep Curve Estimation). First, based on the Zero-DCE model, the U-Net is employed to improve the backbone network DCE-Net for curve estimation, and a coordinate attention mechanism is added to enhance the dark light perception ability of local image areas. Second, the NAF-Net noise removal module is added after the curve estimation backbone network to effectively suppress the noises after low-light enhancement by Zero-DCE. To offset the distortion and overexposure of the enhanced images, the 4-neighborhood calculation method of the spatial consistency loss function is extended to an 8-neighborhood calculation method, enhancing the smoothness of the outputs. Through the ablation experiment on the LOL dataset, the DA-Zero-DCE model, compared to the Zero-DCE model, improves PSNR by 10 dB and SSIM by 0.1, demonstrating its feasibility and effectiveness.