Insect Classification Using Squeeze-and-Excitation and Attention Modules - a Benchmark Study

Yoon Jin Park, Gervase Tuxworth, Jun Zhou

Abstract: Insect recognition at the species level is an active research field with a variety of applications. With the advancement of convolutional neural networks an automatic fine-grained image classifier has displayed encouraging performance. Despite these recent advances, differentiating images at the species level is still a challenge. To address the problems arising from insect-specific peculiarities, this paper presents a novel network that consists of squeeze-and-excitation modules and attention modules, enabling the network to focus on more informative and differentiating features with a limited number of training iterations and a small dataset. The proposed model is trained on an insect dataset collected from Atlas of Living Australia. The results reveal that the integrated model achieves higher accuracy than several alternative methods on the introduced insect dataset.