FasterPest: A Multi-Task Classification Model for Rice Pest Recognition

In the field of precision agriculture and plant protection, this paper proposes a multi-task classification model named FasterPest to improve the accuracy of rice pest identification. Based on the FasterViT network, this model can identify rice pest species and leaf conditions simultaneously. Specifically, the model first sets up two classification heads for multi-task outputs. It then incorporates a feature fusion module that effectively integrates the base model’s features and the classification heads’ output features using attention mechanisms. The fused features are fed back to the classification heads. To address cases where some pest image features are not distinct, the model leverages the correlation between leaf condition and pest species by introducing a relationship matrix, enhancing the model’s capability to identify rice pests. Experimental results demonstrate that the model performs exceptionally well in recognizing 14 classes of rice pests in the IP102 dataset. Compared to the baseline model, FasterPest shows improvements of 6.90% in Accuracy, 5.93% in Recall, and 5.79% in F1 score. The model exhibits outstanding performance in rice pest identification, with significant potential value for practical applications.