Few-shot crack detection based on image processing and improved YOLOv5

The outstanding generalization ability of deep learning methods, such as You Only Look Once Version 5 (YOLOv5), is acquired by training with a large amount of labeled data. However, sufficient data is difficult to be obtained in industrial practices. There are some effective ways for the few-shot object detection such as data augmentation and optimization of the model. This research attempts a method for promoting detection performance of crack detector with only 80 images, that is, running two-type image processing initially, aiming at improving the quality of samples and expanding the quantity of data. All images are sharpened through a Laplacian filter before they are input into the network. Additionally, offline and online data augmentations are used simultaneously for increasing the diversity of the training set. The optimal online data augmentation strategy is obtained in fewer experiments by using a proposed search algorithm. After image processing, improved YOLOv5 is achieved via integrating Convolutional Block Attention Modules and transformer blocks into the YOLOv5l model which can refine the feature maps for prediction. The results show that Recall is improved from 0.560 to 0.742 by 39%, and F^(0.8), a variant of the F1-score for crack detection, is improved by 28%. It can be seen that the preceding procedures successfully improve performance of the crack detector. Also, this research can provide a reference and guide for object detection tasks with small dataset.