M2DETR: A Multi-band Multi-scale Detection Transformer for Pest Detection

Abstract

Accurate pest detection is crucial for ensuring high crop yields. However, existing pest detection models often fail to effectively extract multi-scale features and suppress noise interference from pest images, leading to suboptimal detection performance. To address these limitations, we propose a Multi-band Multi-scale Detection Transformer (M2DETR) model. The core of M2DETR is a novel Multi-band Multi-scale Denoising (MMD) encoder, which integrates Multi-scale Feature Enhancement and Denoising (MFED) modules and Multi-band Multi-scale Downsampling and Denoising (MMDD) modules. Both modules are built upon Cross-scale Convolutional Attention Denoising (CCAD) blocks, designed to filter noise from input feature maps. This enables the MMD encoder to extract high-quality multi-band multi-scale features for the subsequent decoder. Extensive experiments on two public datasets demonstrate the superiority of M2DETR. On the COCO dataset, M2DETR achieves an average precision (AP) of 53.4% and an AP for small objects (AP${}_{\text{S}}$) of 35.6%, surpassing DINO-DETR by 2.5% and 1.0%, respectively. On the IP102 pest dataset, M2DETR outperforms YOLO-Pest by 5.6% in AP50 and exceeds RT-DETR by 1.7% and 1.2% in AP50 and AP, respectively. Moreover, M2DETR exhibits superior noise resistance compared to state-of-the-art models. Our code is available at https://github.com/tanwb/M2DETR-master.