Surface defect detection method for air rudder based on positive samples

In actual industrial applications, the defect detection performance of deep learning models mainly depends on the size and quality of training samples. However, defective samples are difficult to obtain, which greatly limits the application of deep learning-based surface defect detection methods to industrial manufacturing processes. Aiming at solving the problem of insufficient defective samples, a surface defect detection method based on Frequency shift-Convolutional Autoencoder (Fs-CAE) network and Statistical Process Control (SPC) thresholding was proposed. The Fs-CAE network was established by adding frequency shift operation on the basis of the CAE network. The loss of high-frequency information can be prevented through the Fs-CAE network, thereby lowering the interference to defect detection during image reconstruction. The incremental SPC thresholding was introduced to detect defects automatically. The proposed method only needs samples without defects for model training and does not require labels, thus reducing manual labeling time. The surface defect detection method was tested on the air rudder image sets from the image acquisition platform and data augmentation methods. The experimental results indicated that the detection performance of the method proposed in this paper was superior to other surface defect detection methods based on image reconstruction and object detection algorithms. The new method exhibits low false positive rate (FP rate, 0%), low false negative rate (FN rate, 10%), high accuracy (95.19%) and short detection time (0.35 s per image), which shows great potential in practical industrial applications.