Fine-Grained Graptolite Image Classification Based on Multi-Scale Deep Learning

Graptolite is the fossil of the graptolite fauna. Experts can identify the chronological order of strata through graptolite species. Graptolite image classification is much more challenging than traditional fine-grained image classification tasks due to the impacts of geological layer mining extrusion, light and noise. In this paper, we propose a multi-scale deep learning method to tackle these problems. By integrating feature from different scale learning, we can accurately locate the discriminative regions. Training these discriminative part images can further identify the subtle differences, and the proposed model is optimal for adapting to the graptolite classification task. Experimental results on the benchmark graptolite dataset show that our method achieves the state-of-the-art performance.