Visual Attention Focusing on Fine-Grained Foreground and Eliminating Background Bias for Pest Image Identification

Plant diseases and pests caused by harmful insects has always been a significant threat to agricultural and forestry production. In addition, the threat of invasive insects causes damage to local ecosystems with a decrease in biodiversity and even the extinction of some species, seriously harming the local economy. Governments around the world have invested a significant number of efforts in insect detection and control. With the development of AI, automated identification is an irreversible trend to improve efficiency and reduce government input. Recent researches attempt to apply deep learning tools into the detection and identification of insects, but meeting a series of difficulties. Insect identification abstracted to a fine-grained vision classification task provides unique challenges including the small difference between classes and the large difference within a class. In this study, we propose a pest identification model guided by visual attention, designed to address the above challenges. We establish an attention mechanism from these two perspectives, enhancing attention to foreground features by amplifying fine-grained features and eliminating attention to background biases through counterfactual inference. Our approach ultimately achieves a classification accuracy of 74.5% for 102 insect categories on the IP102 dataset, and similarly, achieves an exceptional 99.8% accuracy for 40 insect categories on the D0 dataset. The approach proposed in this study will contribute to the automatic insect detection and identification system in the future as the core technique.