Automatic pavement crack detection is vital for road infrastructure maintenance. However, existing deep learning models fail to effectively balance detection accuracy with model complexity. To tackle this challenge, a novel lightweight multilevel thresholding segmentation crack detection (MTSCrack) model is proposed in this paper. MTSCrack employs a newly devised procedure adopting pre-processing to achieve a lightweight design. Pre-processing is performed by a multilevel thresholding image segmentation method. Multilevel thresholding segmentation decreases the complexity of input data by quantizing an image into K binary planes. To practically perform multilevel thresholding segmentation, a multilevel thresholding segmentation algorithm with low computational cost is proposed. K binary planes are fused by a convolutional network employing encoder-decoder structure to generate a binary crack prediction map as output. Since the complexity of the input data is low, the encoder and the decoder can adopt simple and lightweight architectures. The convolutional network contains only 0.19M parameters. Thorough performance evaluation experiments were performed. 8 was employed for K. Data split ratios of 1.1:1 and 3:1 were employed on Crack500 dataset and DeepCrack dataset, respectively. Image size of 256 × 256 was employed in most of the experiments. No augmentation was employed. Experimental results indicate that the overall performance of MTSCrack outperforms existing lightweight crack detection models and is comparable to heavyweight models. MTSCrack achieves F1-score of 0.820 and 0.756 on DeepCrack and Crack500, respectively. In addition, MTSCrack achieves computational complexity of 2.63G FLOPs and inference speed of 180 FPS on RTX 3090. MTSCrack was deployed on Jetson TX2 and Jetson XAVIER NX and achieved inference speeds of 15 FPS and 35 FPS on Jetson TX2 and Jetson XAVIER NX, respectively.
Zhang et al. (Sun,) studied this question.