Authors - Eric Khang Heng Ooi, Yit Yin Wee Abstract - Printed circuit boards (PCBs) are increasingly dense and visually complex, making reliable component detection important for automated optical inspection and manufacturing quality control. This paper presents a modified R-CNN framework for integrated-circuit (IC) component detection in PCB images. The framework consists of PCB region extraction, colour-guided region proposal generation, Bayesian convolutional neural network (BCNN) feature extraction, and support vector machine (SVM) classification. Candidate regions are produced using colour masks that emphasize dark and silver IC-like regions. Each candidate is resized, represented by the BCNN, and classified by the SVM as IC or background. Experiments on a PCB component dataset show that the proposed method achieves an mAP of 0.613, outperforming R-CNN with Selective Search Fast (0.392) and Selective Search Quality (0.430). YOLO obtains the highest mAP of 0.686; however, the proposed framework remains useful as an interpretable region-based pipeline for PCB inspection.
Open Zoom