Training Noise-Robust Deep Neural Networks for Socket Defect Detection

This paper proposes the development of a noise-robust CNN network for the detection and classification of highly noisy socket image data. Manually inspecting and identifying all surface defects on sockets amidst noise in a manufacturing environment is a significant challenge. Noisy data can greatly impair the performance of systems in computer vision applications. A common approach is sample selection, which involves choosing clean data from a noisy dataset; however, this method is not feasible in this context, where noise is present in all image data. Consequently, we suggest a novel and effective method that, in contrast to most existing methods, involves training a deep learning model with noisy data. Our experiments across multiple benchmarks demonstrate the state-of the-art performance of our method and its enhanced noise-robust capabilities. The proposed model can successfully detect various socket defects in the presence of noise with 96% accuracy, which can help mitigate the costs associated with microprocessor defects on the production floor. We hope this article will assist researchers in selecting new techniques that effectively address surface defect detection.