Improving classification accuracy of a convolutional neural network : case: wood log classification using x-ray images

Abstract

Convolutional neural networks have become a popular way to classify images in a variety of areas, for example, medicine or industry. Convolutional neural networks can assist specialists in giving analyses and increase production performance. This thesis is aimed at assisting with the case of a Finnish industrial sawmill, by applying methods that can improve performance of a convolutional neural network.

The case involves classification of wood logs into “good” and “defective” ones by using their x-ray images. The goal is to improve the classification accuracy of a convolutional neural network. The thesis begins with an introduction, describing the motivation, methodology and research aims in more detail. Following that, a theoretical background is given, where key information about artificial neural networks is described. Next, an extensive literature review is carried out to determine key challenges researchers and practitioners faced in similar fields and with similar problems, and to pinpoint the most common methods to deal with these challenges. The reviewed fields include medicine, industry (most closely related to the case of this thesis) and miscellaneous applications, which dealt with issues like the researched case. Finally, the review is summed up, with the most common method of accuracy improvement being data preprocessing image augmentation and transfer learning.

Data augmentation is applied to the real case convolutional neural network based on the AlexNet architecture. Overall, by applying this method, it was possible to improve the classification accuracy of the neural network from 80.65% to 89.13%, with a baseline dataset size of 306 images. The results concur with previous research (see, e.g., Heidari et al. 2020, X. Li et al. 2022 and J. Li et al. 2021.).