Machine vision inspection for industrial automation
Manufacturing sectors consistently oversee their products and components for accuracy and quality. Identifying defects during production is crucial as it reduces waste of raw materials and time, while also facilitating effective quality control. A client with a heavy industry manufacturing facility sought to enhance their production process by incorporating a tailored machine vision system designed to measure dimensions and identify defects in their manufactured goods. To accomplish this, we needed to analyze the manufacturing process and the products based on various metrics such as length, thickness, and surface texture. The primary objectives of the project were to detect the object of interest from the camera feed and measure it accurately. Since the object did not conform to a standard shape, traditional shape recognition techniques, like Hough Transforms, were inadequate for detection. Thus, we initiated the data acquisition phase by capturing video from the cameras and isolating segments related to the object of interest. This was succeeded by an annotation phase using the CVAT tool to create a segmentation dataset consisting of 5000 images. We implemented a neural network training process based on CNN architectures, including Xception, Inception, and ResNet. Our Receiver Operating Characteristic (RoC) analysis indicated that the ResNet model was performing well regarding both accuracy and processing time. Subsequently, we adapted the ResNet inference program to an edge computing platform, which was integrated with the existing camera module. This platform processed the camera feed in real time, measuring the objects and transmitting the measurements over a network connection.