A machine vision approach using real-time parallel computing for multi-camera quality control systems

Today, thanks to the capable machine learning models introduced into the market and the powerful hardwares that these models can easily work on, machine learning technology has a wide area of use in the industrial field as well as in many areas. One of the important uses of machine learning techniques in this field is computer vision-based quality control, which checks for defects on products. These tasks, which were previously performed by humans and based on their ability to see and make decisions, are now tried to be fulfilled through intelligent systems. For this purpose, in this study, it has been tried to create a model for industrial quality control processes, which is a new and popular subject in the field of machine learning. Artificial neural networks, a sub-field of machine learning, were used for the model to be developed within the scope of the thesis. For this purpose, YOLO, a CNN algorithm, was preferred for the model in order to perform object recognition tasks. The objects recognized by the recognition algorithm and extracted from the image were then subjected to the feature extraction process using the ORB algorithm, and by using the combination of the brute force algorithm and the k-nearest neighbor algorithm, the target object and the source object were compared by making feature matching. In the study, 3 cameras were used in order to obtain simultaneous images from the live environment and with the help of these cameras, images of the target object were taken from different angles and transmitted to the object recognition algorithm. In order to enable the YOLO algorithm to perform the task of simultaneous object recognition, with the help of the CUDA library, the GPU that allows parallel processing was used as an accelerator. Obtained results and findings were given.