Optimizing Automated Detection of Cross-Recessed Screws in Laptops Using a Neural Network

Abstract

This paper investigates varying the operating conditions of a neural network in a robotic system using a low-cost webcam to achieve optimal settings in order to detect crossed-recess screws on laptops, a necessary step in the realization of automated disassembly systems. A study was performed that varied the lighting conditions, velocity, and number of passes the robot made over the laptop, as well as the network size of a YOLO-v5 neural network. The analysis reveals that specific combinations of operating parameters and neural network configurations can significantly improve detection accuracy. Specifically, the best results for the majority of laptops were obtained when the system ran at medium velocity (10 and 15 mm/s), with a light, and the neural network was run with an extra large network. Additionally, the results show that screw characteristics like the screw hole depth, the presence of a taper in the screw hole, screw hole location, and the color difference between the laptop cover and the screw color impact the system’s overall detection rate, with the most important factor being the depth of the screw.