Wood species identification based on an ensemble of deep convolution neural networks
Our paper proposed an ensemble framework of combining three deep convolution neural networks (CNN). This method was inspired by network in network. Transfer learning used to accelerate training and deeper layers of network. Nine different CNN architectures were trained and evaluated in two wood macroscopic images datasets. After two times of 30 epochs training, our proposed network obtained 100% test rate in our dataset, which including 8 kinds of wood species and 918 images. The proposed method achieved 98.81% test recognition rate after three times training with 30 epochs in other dataset, which including 41 kinds of wood species and 11,984 images. Results showed that magnification macroscopic images can be instead of microscopic images in wood species identification, and our proposed ensemble of deep CNN can be used for wood species identification.