Electrical Components Detection in Images with YOLO Model Architectures Using Slicing-Aided Hyper Inference

Athanasios Delis, Antonios Alexiadis, Panayiotis Tsanakas

Abstract: This study tries to solve the problem of accurately identifying electronic parts in pictures of printed circuit boards and circuit diagrams. Traditional object detection methods often fail at this task either because they don't have enough receptive fields or feature resolution. To address this issue, we propose the integration of Slicing-Aided Hyper Inference (SAHI) with cutting-edge detectors. SAHI segments the original images into overlapping $256 \times 256$ slices, facilitating concentrated research on small, highly clustered objects. Three models-YOLOv8, YOLOv11, and RTMDet-were trained on two datasets comprising Printed Circuit Boards (PCBs) and schematics. Experimental findings indicate that SAHI-based inference consistently enhances detection performance, as reflected by an improvement in bounding box mean Average Precision relative to full-image inference. The results indicate that careful slicing and result aggregation can make finding electrical components more accurate by localizing more small components. These models can act as domain-specific visual modules by providing precise component localization at inference time and can be used in downstream tasks and integrated seamlessly in systems such as Large Language Model (LLM) circuit design assistants.

External IDs:dblp:conf/mocast/DelisAT25