VECSNet: A Nondestructive Automatic VCSEL Chip Detection Network With Pixelwise Segmentation

Dark line defects (DLDs) are critical factors that significantly limit the performance of vertical-cavity surface-emitting lasers (VCSELs). Recently, convolutional neural network (CNN)-based methods have shown strong feature extraction capabilities, achieving exceptional performance across various fields. However, these methods still face limitations on the segmentation samples with weak texture, varying shapes and blurred boundary information. To overcome these limitations, a novel segmentation method named VECSNet is proposed in this work. Electroluminescence imaging technology is employed to capture the emission characteristics of VCSELs and develop the corresponding dataset. To improve the extraction of emission features, a parallel dual-encoding structure is designed to capture both spatial and semantic information. Additionally, a feature fusion attention (FFA) block is introduced to effectively fuse features extracted from different branches. Faced with blurred boundary information, a boundary detector is proposed to guide each fusion connection in acquiring boundary feature information and enrich feature representation. Furthermore, to improve segmentation precision for areas with varying shapes, auxiliary logits are introduced to enhance discriminative ability of the network from multiple levels. Experimental results on the VCSEL emission segmentation dataset demonstrate that the proposed method achieves a high Dice score (92.5%) with fewer parameters (6.4M), outperforming other state-of-the-art segmentation approaches.