Comparison of Convolutional Sparse Coding Network and Convolutional Neural Network for Pavement Crack Classification: A Validation Study

Accurate and effective identification of pavement cracks can provide a reference for pavement performance evaluation and prediction. Inspired by recent outstanding performance of convolutional sparse coding theory in various research fields, we envision whether convolutional sparse coding can be beneficial for crack detection in complex environments. Therefore, based on the multi-layer convolutional sparse coding (ML-CSC) model, this paper combines multi-layer iterative soft threshold algorithm (ML-ISTA) and convolutional neural network (CNN) into recurrent neural networks (RNN) to identify crack images across the simulated experiments. And in different noise environments, different training cycles and different epochs, the ML-ISTA is compared with traditional CNN and the layered basis pursuit (LBP), which is another popular algorithm for ML-CSC. Experimental results showed that under the different training conditions with the same parameter setting, the stability and accuracy of the ML-ISTA is better than CNN and LBP. The ML-ISTA can achieve crack identification accuracy of 99.36% efficiently, which demonstrates the effectiveness of convolutional sparse coding in crack detection.