Transfer Learning Approach in Automatic Tropical Wood Recognition System

Abstract

Automatic recognition of tropical wood species is a very challenging task due to the lack of discriminative features among intra wood species and very discriminative features among inter class species. While many conventional pattern recognition algorithms have been implemented and proven to solve wood image classification with 100% accuracy, when using deep learning however, the classification accuracy drops tremendously to only 36.3% due to small number of training samples. Deep learning requires large number of samples in order to work well, unfortunately, wood samples provided by the national forest institute are limited. In this paper, we explore the use of transfer learning in deep neural network for the classification of tropical wood species based on image analysis. Several model of deep learning techniques are tested and results have shown that the classification performance after transfer learning was added reaches 100% accuracy.