A lightweight open-world pest image classifier using ResNet8-based matching network and NT-Xent loss function

Pest control is essential for crop planting as crops are highly susceptible to pest damage. In general, pest recognition is a fundamental element of pest control. Previous works have used computer vision to achieve automatic pest recognition. However, only a few of them have focused on the open-world pest recognition problem. That is, most methods cannot process new pest categories without expensive network retraining. To fill the gap, this paper proposes an open-world pest image classifier based on two observations: (1) convolutional features learned from previous pest classes are generally applicable to new pest categories, and (2) removing fully-connected neural layers allows a deep network to be exempted from model fine-tuning in case of a new class. First, an optimized lightweight ResNet8-based matching network is developed as the image feature extractor, which saves computational resources. To prevent model collapse, the proposed ResNet8-based matching network is trained with the normalized temperature-scaled cross-entropy loss function instead of the triplet loss function. The trained ResNet8-based matching network is then used to compute similarities between support class prototypes and query image representations for the pest classification. Compared with the state -of-the-arts, the proposed method has achieved the highest 40-way 5-shot classification accuracy of 84.29 +/- 0.23% with 14.18 frames per second on the D0 dataset. It is significantly superior to the ResNet12-based baseline. These suggest that the proposed method is a technically feasible solution to the open-world pest recognition problem. The Python code can be accessed at https://github.com/scau-gqw1993.