Learning Causality Representation for Few-shot Pest Recognition

Pest recognition is an effective aid for precisely controlling pests. For pest recognition, there are many types of rare pests with only a few samples, which makes traditional recognition networks based on large-scale supervised data fail to complete this specific task. This can be attributed to the lack of pest samples, as a result of which, network models inevitably associate pests with confounding factors caused by redundant representations such as background. Hence, a causal model is in-troduced to alleviate the confounding problem by learning causal representation in this paper. Concretely, we first build a Causal Structure Model for pest recognition to reveal the confounding factors that affect recognition performance in theory. Then, we design a network based on causal analysis which consists of two modules via two-stage training: 1) An AutoEncoder is built with an unsupervised representation learning approach to extract sufficient and unbiased representations. 2) Afterwards, a Contrastive Masked Frequency Module acts as the intervention operation to obtain causality representations. Experiments on the QFSD and CUB pest datasets demonstrate that the proposed algorithm achieves impressive results and outperforms other competitive pest recognition methods.