-Title: LyFormer: A Lightweight YOLOv8-based Transformer with Four-Stage Preprocessing for Small Semiconductor Component Detection in Smart Manufacturing
-Conference: International Conference on Computer Vision(ICCV 2025), Oct 19 – 23th, 2025, Honolulu, Hawai'i
-Authors: Jinwoo Park, Jeongsuk Ko, Jongpil Jeong
-DOI:
-Conference Link: https://iccv.thecvf.com/

Abstract: With the rapid growth of artificial intelligence (AI) technology, manufacturers are increasingly transitioning to smart factories by replacing traditional rule-based systems. Accurate component counting is especially critical in surface mount technology (SMT) due to its impact on supply chain management and production efficiency. However, smaller enterprises often lack advanced sensors, leading to frequent inaccuracies. To address these challenges, we introduce LyFormer, a specialized small-object detection model built upon YOLOv8. By integrating four preprocessing steps—Label Normalization, Correlation, Local Texture, and Context extraction—and augmenting YOLOv8 with an additional layer, head, and a Transformer module, we enhance the detection of small semiconductor components. Experimental results on X-ray images confirm that LyFormer achieves a mean Average Precision (mAP) of 0.672 (about 0.273 points higher than baseline YOLOv8) and an accuracy of 0.915. These improvements underscore LyFormer’s practical utility for real-time component counting and its potential for seamless integration with Manufacturing Execution Systems (MES).

-Status: Submitted(2025/02/26)