Application of Computer Vision and Neural Network Technology in Intelligent Detection of Road and Bridge Surface Diseases

Existing intelligent detection of road and bridge surface defects lacks high-quality and comprehensive annotated datasets, resulting in inadequate capacity for extension of the model and poor adaptability in complex environments (such as low light, bad weather, etc.), which affects the detection accuracy. This paper proposes a comprehensive method combining deep learning and traditional image processing technology for intelligent detection and classification of road and bridge defects. First, the defect images are manually annotated by Labelmg to construct a multi-category dataset containing five types of defects. Secondly, the improved Inception and ResNet network structures are employed to extract features at many scales, optimize network performance, and reduce model complexity. In view of the problem of limited image quality, histogram equalization and sharpening algorithms are used to enhance the image to improve the recognition of defect features. Additionally, to improve the model's capacity for generalization, the multi-scale HOG feature pyramid captures both the global and local characteristics of cracks. Finally, the improved Inception and ResNet are selected as the backbone feature extraction networks, and compared with traditional CNN, VGG16 and GoogleNet to evaluate the feature extraction ability, detection accuracy and computational efficiency of the models. Results from experiments indicate that this method can achieve efficient detection and classification of defects in complex environments, and provides new ideas and technical support for the identification of surface flaws on roads and bridges.