A Crack Classification Method for Dam Image Based on Double-input Neural Network

Currently, there are many researches on crack classification of underwater dam images, but there are few studies on crack classification of underwater dam images. In recent years, crack classification based on deep learning method has been widely concerned in the field of dam crack classification. However, the general network parameters are huge and the structure is complex. When the number of training samples is too small, it is difficult for deep learning methods to play a powerful role. In this paper, aiming at the underwater dam image, a double-input neural network crack classification method based on feature fusion is designed. The outline of a crack in an image of an underwater dam produces a low-frequency signal, and the texture of the concrete surrounding the crack produces a high-frequency signal. To distinguish the high-frequency and low-frequency signals in the image and make the neural network better learn the characteristics of the cracks in the underwater dam image, this method converts the image into a Fourier frequency domain distribution image for the network to learn. The original image and the frequency domain distribution image are input into two feature extraction modules to extract high-level features respectively, and then fuse the two features to make better use of different features, and finally classify the cracks in the underwater dam image based on the fused features. This paper employs the dataset on underwater dam crack and compares the performance of the method proposed in this paper with that of VGG-16 and AlexNet networks. The findings reveal that the accuracy of the method proposed in this paper surpasses that of VGG-16 and AlexNet networks. Empirical evidence indicates that the approach outlined in this paper has yielded a favorable outcome in the identification of cracks in images of underwater dams. In order to assess the normalization capacity of the model proposed in this study, the Concrete Crack Images for Classification dataset was utilized as input for the model's training, validation, and testing phases. The empirical findings indicate that the model presented in this paper exhibits a strong capacity for generalization.