Optimized VGG Network with Dilated Residual Convolution and Path Enhancement for Crack Image Segmentation

Abstract

Image crack segmentation is a critical task in infrastructure maintenance, as accurate crack detection is essential for structural health monitoring and preventing potential risks. Although traditional Convolutional Neural Networks (CNN) have achieved certain successes in image crack detection, they still exhibit limitations in handling noisy images and detecting fine cracks. This study proposes a VGG network based on dilated residual convolution and path-enhanced optimization to improve the accuracy and efficiency of image crack segmentation. The study utilizes fuzzy morphological filtering for preprocessing, introduces dilated convolution to expand the receptive field, employs residual structures to enhance feature transmission, and incorporates path enhancement modules to boost network performance. The experimental evaluation was conducted using the SDNET2018, METU Dataset, and CFD datasets. The results show that the proposed method achieves a mean intersection over union (MIoU) of 94.60% and a mean accuracy of 96.18%. The improved algorithm demonstrates significant advantages in noise interference handling and detail processing.