Road Image Deblurring with Nonlinear Activation Free Network

Current deblurring methods struggle with real-world scenarios where images are often blurred or noisy, posing significant challenges to existing pavement crack detection techniques. Thus, the aim of this research is to develop and evaluate a novel approach utilizing a nonlinear activation-free network (NAFNet) to address image blurring as a preprocessing step, with the primary goal of improving the reliability and accuracy of pavement crack detection in standard datasets and real-world pavement images under various challenging conditions. The scope of this study is to enhance pavement crack detection by developing a robust and accurate NAFNet designed specifically for road image deblurring, evaluated using standard pavement crack datasets. We adopt NAFNet, which innovatively replaces batch normalization with pixel-level layer normalization and utilizes a U-Net structure with skip connections and optimized the network with SGD (NAFNet-SGD). From the experimental results, quantitatively, the NAFNet-SGD model outperformed the others, achieving the highest PSNR of $\mathbf{3 2. 8 6 4 2}$ and an SSIM of 0.9605, while qualitatively, images processed with NAFNetSGD exhibited the highest quality with superior visual clarity and sharpness. Thus, in conclusion, NAFNet-SGD outperforms other optimizers like Adam and AdamW in terms of both quantitative metrics and visual quality.