Hyperparameter tuning of convolutional neural networks applied to product quality classification

Abstract

The classification of product quality based on their visual aspects is extremely important for many industrial sectors since it contributes to adding value to products by targeting them to the most appropriate consumer market. In general, this task is performed by human specialists, who are subject to several factors that can make this process subjective and costly. Machine Learning (ML) techniques have been successfully used to automate this process. Recently, Convolutional Neural Networks (CNNs) have been extensively investigated for this purpose due to its higher performance compared to other techniques for image classification. On the other hand, just as other ML techniques, CNNs are affected by the choice of their hyperparameter values, which must be previously defined by the users and influences the performance of the induced models. However, there is no guideline on how these values should be set, and most users do by trial-and-error or random search. Therefore, this research project aims to investigate the problem of CNN hyperparameter tuning to improve predictive performance in the visual classification of product quality. Different methods for defining the values of these hyperparameters will be analyzed, such as Sequential Model Bayesian Optimization. The predictive performance of the models generated by the CNNs will be assessed by measures such as balanced accuracy (BAC) and area under the ROC curve and compared with other techniques. (AU)