Automated Defect Detection in Physical Components using Machine Learning

Abstract

It is a crucial part of any manufacturing process, either using manual inspection or using today's modern approaches, to detect the defects at the earlier stages to minimise the risks of failure at later stages. In the early days, manual inspection was prone to many errors, leading to a loss of resources and was very time-consuming. Among the other research areas, it is also an active field of research to achieve the perfect balance between high performance and accuracy in defect detection. ResNet, AlexNet, GoogLeNet, and VGGNet has shown remarkable improvement over old traditional designs in this regard. Image processing and deep learning-based object detection model adopted by Google Cloud Machine Learning Engine were widely used for defect detection and had shown somewhat satisfactory results. In this paper, we proposed a model which is successfully trained on the Google Cloud ML Engine. The results have shown that MobileNet-SSD can automatically detect surface defects more frequently, accurately, and precisely compared to conventional deep learning methods. We have used the pre-trained model of MobileNet V2, which is already trained on lakhs of images and is resource-efficient because it needs small memory setup and lower processing power of the CPU.