A deep learning model for detecting PCB defects based on arbitrary-oriented object detection was proposed, which could simplify the whole process by avoiding image registration and image pre-processing. This model had two new components, including a Rotation Region Proposal Network (RRPN) and a Rotation Region-of-Interest (RRoI) pooling layer. Experiment results validated the effectiveness of the proposed model by achieving 97.2% mAP on perturbated DeepPCB dataset.