STFE-YOLOX: A Small Target Recognition Network for Surface Mount Component Detection

It is of great significances for the accurate detection of surface mount components in the SMT (Surface Mounted Technology). However, due to the small size of the surface mount components and their dense distribution on the PCB (Printed Circuit Board), the accurate and effective detection of them has become a challenging task. This paper proposes an improved small target detection network based on YOLOX (renamed STFE-YOLOX), in order to solve missing and error problems in the surface mount component detection. Firstly, according to the distribution characteristics of surface mount components, a Moudle data enhancement strategy is proposed to improve the original network, which can broaden the range of data and enhance the model’s training effectiveness. Next, a MFFM (multi-feature fusion module) is proposed to fuse features of different levels in the network, which adds a detection branch for improving small target detection capabilities; The experiment results show that mAP
50
(%) and mAP50:95(%) of the proposed network in this paper reach 91.4% and 60.7% respectively on the surface mount component data set, it balances detection accuracy and speed better than the current standard detection models.