RustGAN: A method to generate rusty screw with physicochemical constraints

• Published in: International Conference on Advanced Mechatronic Systems pp. 7 - 12
• Main Authors: Lu, Sicheng, Xu, Chen, Lin, Qiong, Gao, Yicong, He, Zhiliang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 26.11.2024
• Subjects: Assembly; Corrosion; Fasteners; Generative adversarial networks; Machine vision; Maintenance; Mechatronics; Object detection; physicochemical constraints; Recycling; rust screws; RustGAN; Target recognition
• ISSN: 2325-0690
• DOI: 10.1109/ICAMechS63130.2024.10818721

As a crucial and pervasive component in industrial processes, screws are integral to the assembly, maintenance, and recycling of products. In light of the advancements in automation technology, it has become imperative to leverage vision systems for the identification and execution of automated disassembly and assembly of screws. However, the working environment of screws is complex and variable, and corrosion is one of the main causes of damage to screws. Therefore, it is essential that vision systems are able to accurately identify different types and degrees of corrosion in order that appropriate treatment measures can be taken. However, the current challenge is that it is more difficult to obtain samples of screws exhibiting various corrosion, which directly constrains the size of the dataset and, consequently, affects the accuracy and generalizability of the recognition algorithm. In this paper, we propose a Rust Generative Adversarial Network (RustGAN) approach with physicochemical constraints for the generation of images of screws exhibiting different types and degrees of rust. In particular, we initially delineate three screw rusting base models in practical scenarios and subsequently calculate the corresponding formulas under physicochemical constraints. Concurrently, we devise a novel screw rusting detail feature loss function based on the formulas and integrate it into the generative adversarial network to generate high-quality and diverse screw rusting images. The efficacy of the RustGAN proposed in this study has been demonstrated through experimentation, whereby the dataset of screws exhibiting various degrees of rust has been expanded, and the performance of the VGG16 target detection model on the task of recognizing rusty screws has been markedly improved.