Fine-Grained Butterfly Recognition via Peer Learning Network with Distribution-Aware Penalty Mechanism.

Abstract

Simple SummaryAutomatic species recognition, such as butterflies or other insects, plays a crucial role in intelligent agricultural production management and the study of species diver-sity. However, the quite diverse and subtle interspecific differences and the long-tailed distribution of sample data in fine-grained species recognition are insufficient to learn robust feature representation and alleviate the bias and variance problems of the long-tailed classifier on insect recognition. The objective of this study is to develop a peer learning network with a distribution-aware penalty mechanism proposed to learn discriminative feature representation and mitigate the bias and variance problems in the long-tailed distribution. The results of various contrast experiments on collecting the butterfly-914 dataset show that the proposed PLN-DPM has a higher Rank-1 ac-curacy rate (86.2% on the butterfly dataset and 73.51% on the IP102 dataset). Addi-tionally, we deployed the PLN-DPM model on the smartphone app for butterfly recognition in a real-life environment.Automatic species recognition plays a key role in intelligent agricultural production management and the study of species diversity. However, fine-grained species recognition is a challenging task due to the quite diverse and subtle interclass differences among species and the long-tailed distribution of sample data. In this work, a peer learning network with a distribution-aware penalty mechanism is proposed to address these challenges. Specifically, the proposed method employs the two-stream ResNeSt-50 as the backbone to obtain the initial predicted results. Then, the samples, which are selected from the instances with the same predicted labels by knowledge exchange strategy, are utilized to update the model parameters via the distribution-aware penalty mechanism to mitigate the bias and variance problems in the long-tailed distribution. By performing such adaptive interactive learning, the proposed method can effectively achieve improved recognition accuracy for head classes in long-tailed data and alleviate the adverse effect of many head samples relative to a few samples of the tail classes. To evaluate the proposed method, we construct a large-scale butterfly dataset (named Butterfly-914) that contains approximately 72,152 images belonging to 914 species and at least 20 images for each category. Exhaustive experiments are conducted to validate the efficiency of the proposed method from several perspectives. Moreover, the superior Top-1 accuracy rate (86.2%) achieved on the butterfly dataset demonstrates that the proposed method can be widely used for agricultural species identification and insect monitoring.