Comparative Performance Evaluation of One-Stage and Two-Stage Object Detectors for Screw Head Detection and Classification in Disassembly Processes

The process of manually detecting and removing screws during disassembly can be time-consuming and demanding. Automated systems utilizing imaging and deep learning can rapidly and accurately identify screws based on features like head shape and size. Different deep learning methods have varying screw detection capabilities. Choosing the right method significantly impacts system effectiveness, robustness and efficiency. Several generic deep learning architectures have been proposed to address this problem. However, there is little guidance on their suitability for specific scenarios. This paper aims to provide qualified guidance for selecting suitable deep learning methods for screw head detection. The proposed approach involves researching existing object detection methods and choosing two representatives - one employing a one-stage approach, the other using two-stage. These selected methods will be implemented identically and their performance evaluated via rigorous analysis and comparison. The dataset comprises synthetic and real images of 6 screw types across 3 products. YOLOv5 was selected for one-stage detectors, Faster R-CNN for two-stage. Results show the one-stage YOLOv5 generally outperforms Faster R-CNN in precision, recall, speed and training times, while Faster R-CNN delivered superior recall with smaller training datasets.