A point-based method for identification and counting of tiny object insects in cotton fields

Monitoring of crop pests in the field can be achieved by using sticky traps that capture pests. However, due to the small size and high density of the captured pests, conventional object detection methods relying on bounding boxes struggle to accurately identify and count pests, as they are highly sensitive to positional deviations. Therefore, we propose a novel point framework for multi-species insect identification and counting, termed MS-P2P, which is free from the limitation of Bounding box. Specifically, we employ the lightweight object detection network YOLOv7-tiny for feature extraction and incorporate a lightweight attention detection head (LAHead) for point coordinate regression and insect classification. The LAHead enhances the model’s sensitivity to subtle insect features in complex environments. Additionally, we utilize point proposal prediction and the Hungarian matching algorithm to achieve one-to-one matching of optimal prediction points for targets, which simplifies post-processing methods significantly. Finally, we introduce SmoothL1 Loss and Focal Loss to address the issues of matching instability and class imbalance in the point estimation strategy, respectively. Extensive experiments on the self-built NSC dataset and the publicly available YST dataset have demonstrated the effectiveness of our designed MS-P2P. In particular, on our self-built dataset of 9 insect species, the overall counting metrics achieved a MAE of 18.9 and a RMSE of 28.8. The combined localization and counting metric, nAP0.5, reached 86.4%. Compared with other state-of-the-art algorithms, MS-P2P achieved the best overall results in both localization and counting metrics.