Matching veneer sheets images with Siamese neural networks

Abstract

This thesis investigates the use of Siamese neural networks in matching digital veneer sheet images. It reviews the literature on the present use of SNNs in image analysis and their effectiveness in capturing similarity. The thesis then presents an SNN model based on the architecture proposed by Jalonen et al. (2021), using the publicly available Veneer21 dataset (Jalonen, 2021), which consists of 2579 pairs of Dry and Wet veneer sheets. Compared to the benchmark study by Jalonen et al. (2021), which employed an identical dataset and SNN model, this thesis study realized an improved accuracy of 100%, surpassing the previous study’s 84.93% with 50 pairs of veneer sheets. This improvement is primarily due to the augmentation of the dataset in training and a reduction in dense layers in the architecture that had previously resulted in overfitting. Results from this study will be vital in the development and further research of product tracking, especially in the wood industry.