This work focuses on cost reduction methods for forest species recognition systems. Current state-of-the-art shows that the accuracy of these systems have increased considerably in the past years, but the cost in time to perform the recognition of input samples has also increased proportionally. For this reason, in this work we focus on investigating methods for cost reduction locally (at either feature extraction or classification level individually) and globally (at both levels combined), and evaluate two main aspects: 1) the impact in cost reduction, given the proposed measures for it; and 2) the impact in recognition accuracy. The experimental evaluation conducted on two forest species datasets demonstrated that, with global cost reduction, the cost of the system can be reduced to less than 1/20 and recognition rates that are better than those of the original system can be achieved.

Multi-scale Forest Species Recognition Systems for Reduced Cost