Bridge defect detection using small sample data with deep learning and Hyperspectral imaging

The visual sensing method is an effective way to address long-term health monitoring of bridges. However, bridge defect detection based on visible light imaging mainly relies on grayscale and regional edge gradient information, which brings challenges such as limited information dimensions and complex background. This paper introduces a bridge defect detection method that leverages hyperspectral imaging, utilizing the unique integration of spectral and spatial information. Also a convolutional neural network algorithm with dual branches and dense blocks for spectral feature extraction is developed. This framework includes spectral and spatial branches, which independently extract respective features in order to minimize mutual interference. Compared with the support vector machine and traditional deep learning algorithms, the proposed method attains an overall model prediction accuracy(OA) of 98.57 %, an average accuracy (AA) of 98.16 %, and a Kappa coefficient of 0.9814, representing the best classification performance on small sample datasets.