Manufacturing industries continuously monitor their products and artifacts for their precision and quality. It is imperative that they detect defects during their manufacturing process as this eliminates wastage in terms of raw material and the time to produce the same. It also paves way for good quality control during the production process. The client had a heavy industry manufacturing plant and they wanted to improve their manufacturing process by adding a customised machine vision system. They wanted us to build a system that can measure the dimensions and detect defects on the manufactured artifacts. In order to achieve this, we were required to study the manufacturing process and product based on their metrics of manufacturing, namely length, thickness, surface texture etc.
The key goals of the project are to identify the object of interest on the camera feed and to measure the object. Since the object of interest is not a standard shape, conventional shape recognition methods like Hough Transforms are not suitable for object detection. So we started the data acquisition phase of the project by collecting the video recording from the camera feeds and trim the video segments pertaining to the object of interest. This was then followed by an object annotation phase with the CVAT tool to generate a segmentation dataset of 5000 images. We implemented a neural network training process derived from the CNN architectures including the Xception, Inception and ResNet. From our RoC studies, it was observed that the ResNet model was performing adequately in terms of both accuracy and execution time. We then ported the ResNet inference program to an edge computing platform that interfaces along with the existing camera module. The edge computing platform read the camera feed and performed real time object measurement and reported the measurements over a network connection.
We are able to introduce a novel object measurement process in their manufacturing pipeline which was able to detect manufacturing defects, thereby eliminating wastes due to defective product discards. This project was developed in a time frame of 24 weeks. The ResNet based object detection model was more than 95% accurate in its predictions.