CRA-Net: A channel recalibration feature pyramid network for detecting small pests

There are multiple categories of agricultural pests, which poses great challenges to accurate pest recognition. Deep convolutional neural networks (DCNNs) are effective in pest detection due to their powerful feature extraction capabilities. However, for small agricultural pests with few inter-class physical variations, the DCNNs extract fewer effective features, and thus perform poorly. To address this problem, we propose a CRA-Net, which includes a channel recalibration feature pyramid network (CRFPN) and an adaptive anchor (AA) module. CRFPN can capture discriminative features, which significantly improves recognition accuracy and localization with regard to small pests, while the AA module can correct the inefficient matching of anchor and ground truth boxes. To evaluate the performance of the proposed method, several experiments were conducted using our constructed large-scale, multi-category pest dataset. These results demonstrate that our method achieves 67.9% average precision (AP), outperforming other state-of-the-art methods.

Automated summary

This study introduces a new method for agricultural pest recognition, CRA-Net, which enhances the accuracy and localization for small pests by incorporating CRFPN and AA modules. Experimental results demonstrate that the proposed method achieves a high average precision.