Lightweight Multi-Attention Network for SMT Dispensing Electronic Mount Components Identification

During manufacturing process of electronic products, surface mounted technology component dispensing is a key technology in the process of chip production, which affects the quality of products. However, majority dispensing methods utilize rule-based methods, which are not robust to different styles of backgrounds and need to modify parameters when they face different templates. To address the above issues, a YOLOv5-based lightweight multi-attention detection network is proposed for SMT dispensing electronic mount components identification, in which cross and shuffle attention and ghost and multi-attention modules are designed to reduce computational complexity and locate dispensing components rapidly. Moreover, the SMT intelligent dispensing pipeline is realized and experiments on the self-constructed dispensing dataset and experiments show that YOLOv5-Light reaches the best frames 17 FPS on Jetson Nano and satisfied accuracy with mAP@.5 of 99.5% and mAP@.75 of 99.4%. Also, the inference speed is improved from 7.3 ms to 1.0 ms while the space complexity is improved from 10.2 MB to 6.2 MB, which indicates that the proposed dispensing system could implement fast detection under minimizing accuracy loss.