Lightweight detector based on knowledge distillation for magnetic particle inspection of forgings

Magnetic particle inspection (MPI) is a widely-used technique within the realm of non-destructive testing (NDT), aimed at detecting defects located on or just beneath the surface of ferromagnetic parts. Deep learning-based object detection methods have demonstrated the ability to quickly and accurately identify targets within complex backgrounds, a quality well-aligned with the demands of crack detection using magnetic particles in the context of fluorescent MPI. However, current research notably lacks studies that address the challenges of balancing efficiency and accuracy when handling a substantial volume of images, particularly those requiring detection across the entire surface of components. In this paper, a lightweight detector designed for industrial applications of crack detection in forgings, utilizing knowledge distillation, is proposed. The proposed detector is built upon the framework of the SSD (Single Shot MultiBox Detector), with its backbone engineered using depth-wise separable convolution and inverse residuals. This approach effectively reduces the network size, resulting in a lightweight model that maintains high accuracy through knowledge distillation. Additionally, to further enhance the network’s precision, an enhanced loss function and an attention module are introduced. The effectiveness of this methodology is assessed using a dataset of forging crack magnetic particle indications collected through a MPI platform. Compared to the original SSD, the experimental results demonstrate that our method achieves an improved mean average precision (mAP) of 96.57%, which is 5% higher. Moreover, the model size has been reduced by 53%, and the detection speed has increased to 35.25 FPS (frames per second), representing a 15 FPS improvement over the original SSD’s 20.57 FPS. These results suggest promising applications for the proposed detector in low-computing-power and memory-constrained devices, particularly embedded systems.