Fossil Brachiopod identification using a new deep convolutional neural network

Abstract The identification of brachiopods requires specialist knowledge held by a limited number of researchers and is very time-consuming. The new technique of deep learning by artificial intelligence offers promising tools to break these shackles to develop computer automatic identification. However, we found that the traditional convolution neural network is not sufficient to automatically identify brachiopod species. Thus, we propose a new tailored Transpose Convolutional Neural Network (TCNN) in order to automatically identify brachiopod fossils with high efficiency. In this network, we add an “upsampling” Transpose Convolutional Layer and synthesize the data of this layer with the data of a Convolutional Layer to fully mix the small and large scales features extracted by the neural network. Compared with the traditional Convolution Neural Network (CNN), the Transpose Convolutional Neural Network (TCNN) can achieve a high identification accuracy using a smaller training data set of images of brachiopods. Results from this study show that the TCNN can achieve 98%, 98% and 97% identification accuracy respectively, with training data sets of 400 images of 3 species, 484 images of 4 species and 630 images of 5 species. In contrast, the traditional CNN can achieve only a low identification accuracy (67%) with 400 images of 3 species and requires 3000 images per 3 species to achieve a 95% identification accuracy. For most of brachiopod species, it is almost an impossible task to collected thousands of samples and as more brachiopod species are fitted into automatic identification, it is significant to have a reliable network which can achieve high accuracy on a small data set. In summary, the TCNN is a more efficient neural network that could be better applied to automatically identify brachiopod fossils.