Comparing Object Recognition Models and Studying Hyperparameter Selection for the Detection of Bolts.

The commonly-used method of bolting, used to secure parts of apparatus together, relies on the bolts having a sufficient preload force in order to the ensure mechanical strength. Failing to secure bolted connections to a suitable torque rating can have dangerous consequences. As part of a wider system that might monitor the integrity of bolted connections using artificial intelligence techniques such as machine learning, it is necessary to first identify and isolate the location of the bolt. In this study, we make use of several contemporary machine learning-based object detection algorithms to address the problem of bolt recognition. We use the latest version of You Only Look Once (YOLO) and compare it with algorithms RetinaNet and Faster R-CNN. In doing so, we determine the optimum learning rate for use with a given dataset and make a comparison showing how this particular hyperparameter has a considerable effect on the accuracy of the trained model. We also observe the accuracy levels achievable using training data that has been lowered in resolution and had augmentation applied to simulate camera blurring and variable lighting conditions. We find that YOLO can achieve a test mean average precision of 71% on this data.