基于GAN数据增强在航空旋转部件缺陷检测中的应用OA

A defect sample generation method based on the"repair-generation"mechanism is proposed to ad-dress the problem of insufficient and imbalanced defect samples in aviation rotating components.The large mask inpainting(LaMa)model is utilized to perform semantic completion on the original defect images,defect repair image pairs with consistent structures is constructed to enhance the expression consistency of the training samples.At the same time,a mask-guided spatial-frequency feature fusion generative adversarial network(Mask-FFTGAN)is designed to improve the modeling ability of the model for the morphology,boundaries,and local structure of defect areas.The loss function combines conditional adversarial loss and pixel reconstruction loss to achieve balance between the image realism and structural consistency.Experiments on the AeBAD-S dataset show that this method achieves a minimum Fréchet inception distance(FID)value of 10.22 and a learned perceptual image patch similarity(LPIPS)value of less than 0.1 in the task of generating local defect regions.In the defect detection task,the average precision,recall,and F1 score are improved by 10.2%,8.6%,and 9.3%,respectively,through data augmentation.The results show that this method can effectively al-leviate the problem of insufficient samples,improve downstream detection performance,and provide a feasible solution for the generation and enhancement of high-quality defect images in the industrial field.