Neuronal Networks for Visual Inspection of Assembly Completeness and Correctness in Manufacturing

The conformity of the quality with the desired specifications in production must be controlled quickly, reliably and accurately. Cost reduction and efficiency studies in production quality control stages are of great importance today. For this reason, non-human and intelligent automated systems are the main research subjects as a solution method in quality control stages. In this study, the final visual inspection of fastening elements of an industrial product is addressed. The inspection of connection elements, such as screws, as one of these quality control stages, is presented through a framework utilizing a camera and learnable neural network, replacing human-eye control. Fasteners can be counted as small objects in the images obtained. Therefore, in this study, object detectors based on different CNN backbones (ResNet 50–101) and proposals are discussed and their performance in detecting these small objects is compared to achieve the high detection speed, accuracy and reliability. To address the challenges at an industrial level for object detection methods, a non-processed image dataset has been created. This dataset aims to represent various lighting conditions, including dark-bright fields and diffuse reflection, as well as occlusion and restricted camera angles. During the training phase, hyperparameter-tuning optimization of deep networks such as YOLOv8, Faster-RCNN with ResNet50&101 and lastly Sparse-RCNN with a different set of learned object proposals is evaluated, which can be most suitable for the detection of screw connection. Experimental results show that the pretrained Faster-RCNN and Sparse RCNN has over the % 85 success rate of detection of small objects in an industrial environment.