An Application of Semi-Supervised Learning to Convolutional Neural Networks for Surface Roughness Classification

Knowledge about the quality of machine-produced products is crucial for manufacturers. Since the quality of certain parts cannot be determined non-destructively using conventional inspection methods, the demand for nondestructive quality control solutions is continuously rising. Due to recent advancements in deep learning, these methods can now be utilized for non-destructive quality inspection. Specifically, Convolutional Neural Networks have demonstrated exceptional performance in image classification tasks. Therefore, we use a Convolutional Neural Network to evaluate the quality of fine-ground chef knife surfaces by predicting the arithmetic average roughness of the knives and classifying them according to the manufacturer's specification limits. Because Convolutional Neural Networks require a substantial amount of labeled data for training, this technique demands significant time and skilled personnel to annotate the instances. To address this issue, the method of Semi-Supervised Learning is applied. This approach serves as a compromise between the established machine learning methods of supervised and unsupervised learning by utilizing a small labeled dataset alongside a larger unlabeled dataset. The methodology aims to learn from unlabeled data while keeping the labeling costs as low as possible. In this application, we employ the Semi-Supervised Learning method called top-k pseudo-labeling, where the k most confident predictions of the model are assigned the model's predicted labels. Furthermore, we determine the optimal hyperparameters for the proposed algorithm and evaluate its performance from both an inductive and transductive perspective. Additionally, the limitations of the proposed algorithm are discussed, along with explanations for these constraints. Lastly, possible improvements and avenues for further investigation are presented.

Keywords: Semi-Supervised Learning, Convolutional Neural Network, self-training, pseudo labeling, production, Deep Learning.