Pest-PVT: A model for multi-class and dense pest detection and counting in field-scale environments

• The Pest-PVT framework addresses challenges in the Pest24 dataset, including small pest targets, high similarity, and dense distribution. • Pest-PVT uses an anchor-free FCOS method to improve small object detection and integrates ATSS to reduce sample imbalance bias. • Dynamic Heads tackle pests of different scales and spatial changes, while Shunted Self-Attention boosts multi-scale feature extraction and lowers memory use. Field-scale pest monitoring is crucial for evaluating insect infestations in agricultural environments. The Pest24 dataset presents unique challenges due to the small pest target sizes, high target similarity, and dense pest distribution. To address these challenges, we propose Pest-PVT, a comprehensive framework based on the Pyramid Vision Transformer v2 (PVTv2) network model. Pest-PVT adopts an anchor-free approach using Fully Convolutional One-Stage Object Detection (FCOS) to enhance small object detection and incorporates Adaptive Training Sample Selection (ATSS) to mitigate sample imbalance bias. Dynamic Heads (DyHead) are employed to handle pests of varying scales and spatial changes, while Shunted Self-Attention (SSA) enhances multi-scale feature capture and reduces memory consumption. Compared to 23 other mainstream detection models and previously published works, our proposed Pest-PVT outperforms state-of-the-art detectors on pest detection datasets, achieving the outperformed scores in detection evaluation metrics such as mAP, Precision, Recall, and F1-Score, reaching 77.2 %, 78.42 %, 81.27 %, and 0.80, respectively. With a parameter size of only 24.74 M, Pest-PVT is suitable for integration into edge devices with limited resources. This work represents a significant contribution to the field of field-scale pest monitoring and addresses the specific challenges posed by the Pest24 dataset. Our code is made available at https://github.com/jlauwcj/pest-pvt .