Abstract: The convolutional neural network (CNN) algorithm is commonly used in automated crack detection, but its current version is too complicated involving many training parameters, high equipment configuration requirements, and low detection real-time performance. This paper develops a lightweight CNN method (LFNet). This simplified version of CNN reduces the number of training parameters, and then analyzes and extracts cracking features from the images of cracked concrete through a threshold division weight method based on Gaussian gradient change. Finally, it calculates the crack width using a Euclidean distance algorithm. Comparison with experimental results shows LFNet is better than previous methods of classical convolutional neural network and achieves an accuracy, recall and F1 value of 97.9%, 98.3% and 98.1% respectively. Its calculation errors of characteristic crack widths can be controlled within a range of 0.5 mm.

Key words: crack detection, image processing, full convolution neural network, light-weight, threshold weight separation method