Results and Discussions The experiments are conducted on the Google Colab with a deep learning framework of Pytorch 1.7 and a Tesla T4 GPU with 16 GB RAM. In this paper, precision, recall, mean average precision (mAP), and frames rate are taken as model evaluation indicators. In the first comparison experiment, the loss function of the model with the proposed transfer learning method decreases more rapidly and the model converges more stably (Fig. 6). The testing results show that the recall and mAP of the model obtained by the proposed method are improved by 16.5% and 8.9%, respectively (Table 3), proving the effectiveness of the proposed method. In the second comparison experiment, the ASFF-YOLO v5 model has a higher mAP than the baseline, while the loss function decreases faster and the value of the loss function after stability is smaller (Fig. 7). The testing results show that although adaptive spatial feature fusion module increases floating point operations (FLOPs) and leads to a decrease in FPS, the recall of ASFF-YOLO v5 is increased by 11.7%, which effectively overcomes the defect miss detection; the mAP value is increased by 6.2% (Table 4), and the model performance is significantly improved in the recognition and localization of crack defects.1) Through the projection method of target domain feature space, images with similar defect features are selected from the source domain for pre-training the YOLO v5 defect detection model. Then target domain images are adopted to fine-tune the pre-trained model to realize knowledge transfer between similar domains, which solves the problem of insufficient defect samples during model training. The mAP of the transfer learning model is improved by 8.9%, which proves the feasibility of the proposed method.2) The capability of the YOLO v5 model to fuse multi-scale features has been enhanced by the introduction of ASFF. In the comparison experiments, the recall of ASFF-YOLO v5 is improved by 11.7%, which effectively overcomes the miss detection and verifies the accuracy of the method.However, when the crack defect is not obvious or generated in the complex infrared background, the false detection of individual defect sample types may occur. In the future, we will try to introduce attention mechanism to suppress background noise and improve the accuracy of defect detection.

Many key components of mechanical equipment are made of metallic materials. During long-term service, cracks, scratches, pits, and other damages may occur on the surfaces or inside metallic materials. Under the action of the external environment and stress, the damages easily extend to the surrounding area, resulting in more harmful defects with more complicated structures. Therefore, the research on nondestructive testing for steel materials is of great significance to improve the reliability of equipment and prevent catastrophic accidents. Pulsed eddy current thermography has been successfully applied as a visual nondestructive testing method for defect detection in steel materials. The principle is that defects in steel materials affect the distribution of induced eddy currents, which in turn causes changes in the temperature field distribution. However, to ensure the safe operation of the equipment, the actual production often follows a regular cycle of equipment maintenance and component replacement. As a result, a few defect images are collected by pulsed eddy current thermography, and then the defect detection model constructed by the small number of samples suffers from inadequate training, insufficient model generalization, and low defect detection accuracy.

In this study, the construction of the defect detection model based on deep transfer learning is proposed. First, a typical defect sample database is formed by selecting part of infrared images in the target domain, and the target domain feature space is constructed by extracting the defect features through non-negative matrix factorization. Then, the source domain defect images are projected into the target domain feature space, and the images with similar defect features are selected by cosine similarity. In addition, the obtained source domain images are used for pre-training the YOLO v5 defect detection model, and the model weight parameters are transferred to the target domain to realize knowledge transfer in similar domains. Finally, the adaptively spatial feature fusion (ASFF) module is introduced to the YOLO v5 algorithm, and the ASFF-YOLO v5 model is fine-tuned with the training set samples in the target domain and validated with the test set samples to obtain the final crack defect detection model in the target domain.

In this study, a deep transfer learning method for crack defect detection is proposed. Under the experimental platform of this paper, for defect images with a resolution of 320 pixel×240 pixel, the mAP of this model reaches 98.6% and the detection speed is 46 frame/s, which provides references for the development of pulsed eddy current thermography technology toward high efficiency and visualization. The major contributions of this study are summarized as follows.