A Method for Small-Sample Data Augmentation of Weld Seam Radiographs Based on Deep Convolutional Generative Adversarial Networks

Addressing the core issues of scarce small defect samples (cracks) in weld radiographic images, insufficient fidelity of defect details in existing generation methods, and the tendency to lose fine-grained features in full-image generation, this paper proposes an improved DCGAN-based data enhancement method for weld radiographic images, combining local generation with background fusion. This method constructs pre- and post-processing algorithms adapted to the characteristics of industrial radiographic films, reducing the learning difficulty of the network through local cropping and contrast normalization; optimizes the DCGAN fully convolutional network architecture, combining gray-level distribution constraints and defect feature fidelity constraints to achieve high-fidelity generation of local crack samples; and generates complete weld images conforming to DS-level industrial standards through background fusion and inverse gray-level normalization. Comparative and ablation experiments were conducted based on a weld crack dataset conforming to GB/T 3323-2019 standard. Quantitative verification of the enhancement effect was completed using YOLOv8 as the baseline model. Results show that the FID value of the samples generated by this method is as low as 12.36, and the SSIM value reaches 0.947. It can improve the detection model mAP@0.5 to 96.32%, an improvement of 21.47 percentage points compared to the original small-sample scheme. It possesses excellent adaptability to industrial scenarios and can provide reliable technical support for intelligent detection of small-sample defects in the field of industrial radiographic testing.