Abstract

Controllable Electrolyte Distribution Electrochemical Machining (abbreviated as CED-ECM) is a novel ECM method for shallow groove (with a depth of about several μm to tens μm) fabrication on precision metal surfaces. Rapid and accurate detection of CED-ECM-processed grooves is essential for quality control and automated production cycles. However, these grooves exhibit minute scale, irregular morphology, and low contrast against the metallic background, while grayscale microscopic imaging provides only single-channel information. To address these challenges, this research proposes a lightweight instance segmentation network based on pseudo-multimodal wavelet network (abbreviated as PMW-YOLO). First, a pseudo-multimodal channel fusion strategy expands single grayscale images into three complementary channels: original grayscale, CLAHE-enhanced grayscale, and multiscale Sobel gradients. This design explicitly injects illumination robustness and edge priors without additional acquisition cost. Second, a Discrete Wavelet Transform-based downsampling module, termed DWTDown, is integrated into the backbone to preserve high-frequency edge details while reducing model parameters and GFLOPs. Ablation studies further investigate the contributions of an Efficient Multiscale Attention module, a boundary-aware mask loss, and data-centric augmentation strategies. Experiments on an in-house CED-ECM dataset validate the effectiveness of PMW-YOLO for automated groove inspection.

IPC Classification

Keywords