Few-Shot Object Detection for Pest Insects via Features Aggregation and Contrastive Learning

Shuqian He^1*Biao Jin¹Xuechao Sun¹Wenjuan Jiang^1*Fenglin Gu²Fenglin Gu³

• ¹海南师范大学, 海南 海口, China
• ²浙江大学, 杭州，浙江, China
• ³Chinese Academy of Tropical Agricultural Sciences, 万宁 海南, China

Accurate detection of pest insects is crucial for effective agricultural pest control and improved crop yields. However, pest detection faces significant challenges due to the large number of species, considerable individual variations, limited labeled data, and environmental factors, causing traditional methods to perform poorly, especially in small-sample and multi-object detection scenarios. To address these issues, we propose a few-shot pest detection approach based on feature aggregation and supervised contrastive learning (SCL), built upon the Faster R-CNN framework. Our method enhances the feature representation capabilities for small-sample targets by effectively utilizing rich features from a support set. First, multi-scale feature extraction is conducted on both support and query sets using a Feature Pyramid Network (FPN), creating multi-scale feature maps from different backbone levels. Region of Interest (RoI) Align extracts fixed-size feature representations from each RoI. Next, a Feature Aggregation Module (FAM) with an attention mechanism models the relationships between support and query features by projecting them into query, key, and value spaces to calculate attention weights. Aggregated features are obtained through weighted summation, significantly enhancing query set feature expressiveness. To further improve discriminative ability and alleviate misclassification, SCL is employed to increase intra-class similarity and inter-class dissimilarity. Enhanced query features are mapped into a normalized contrastive learning space via a projection head, and positive and negative sample pairs are constructed based on label information. Supervised contrastive loss clusters similar samples while separating dissimilar ones. To mitigate sample imbalance issues common in contrastive learning, class weights are integrated into the loss function, assigning higher importance to minority classes. Additionally, Focal Loss is utilized in classification to emphasize challenging samples. Through multi-task learning, tasks of localization, classification, feature aggregation, and contrastive learning are jointly optimized via a comprehensive loss function design. Experimental evaluations demonstrate excellent performance in pest detection tasks, particularly in multi-object and minority-class detection scenarios. Ablation studies confirm significant contributions from each proposed module. In conclusion, our proposed method effectively tackles critical challenges in few-shot and multi-object pest detection, presenting substantial practical value for improving agricultural pest management and crop productivity.