Adaptive weight optimization with large pretraining for pest detection

Abstract

Frequent infestations by agricultural pests reduce crop production and significantly affect economic efficiency. Therefore, timely and effective pest control is critical to improving productivity and facilitate environmental protection. Herein, we propose an adaptive weight optimization method based on transfer learning for multimodal pest detection. This approach utilizes pretrained model parameters from public datasets to extract features and enhance cross-modal feature from text and images. Accurate pest recognition and localization are achieved through an adaptive loss function, which optimizes the model’s performance across multiple tasks. In tests conducted on IP102 (36 species) and Pest24 (24 species), which are major agricultural pest datasets, the proposed model achieves average precisions of 65.8% and 76.3% at 50% Intersection over Union (IoU) threshold, respectively. By doing so, our model outperforms existing state-of-the-art methods despite using only 30 training cycles. These results highlight the significant practical value of multimodal pest detection methods in enhancing the efficiency and accuracy of agricultural pest identification. The code and dataset are available at https://github.com/Healer-ML/Pest-Detection.