A novel few-shot self-supervised contrastive framework for UAV-based rare pest detection

Résumé fourni par la source

Agricultural pest detection is a highly important problem due to pest rarity, visual similarity, and heterogeneity, causing extreme class imbalance and low generalization in classical supervised models. Current CNN-based and transformer-based approaches are not very accurate but need large annotated data, which cannot be applicable to rare types of pests. Newer self-supervised and contrastive learning methods like SimCLR, BYOL, and MoCo-v3 are more label-efficient, but continue to suffer from high intra-class variations, long-tailed distributions and weak domain generalization in agricultural applications. To overcome these weaknesses, the research paper proposes a label-efficient framework to detect rare pests using few-shot self-supervised contrastive learning. The suggested approach combines synthetic augmentation using Mixup and StyleGAN2, contrastive pretraining on the basis of the SSL, pseudo-labeling using cluster, and few-shot fine-tuning on the lightweight backbones of ResNet and MobileNetV2. It is innovative in the synthetics of rare-class expansion combined with clustered pseudo-supervision and diffusion-inspired augmentation to produce strong and domain adaptive representations. Experiments with the IP102 dataset show that the proposed approach is label efficient (by 12–15 percent), has a precision of up to 96.8 percent and an F1-score of 95.2 percent and performs better than current supervised and SSL baselines. In general, the discussion shows that synthetic augmentation combined with contrastive pretraining is an effective method of detecting the rare pests in a real agricultural setting