key-fg DETR based camouflaged locust objects in complex fields

Dongmei Chen¹Peipei Cao¹Zhihua Diao²Yingying Dong³Jingcheng Zhang^1*

• ¹Hangzhou Dianzi University, Hangzhou, China
• ²Zhengzhou University of Light Industry, Zhengzhou, Henan Province, China
• ³Aerospace Information Research Institute, Chinese Academy of Sciences (CAS), Beijing, Beijing, China

In real agricultural environments, many pests camouflage themselves within complex backgrounds, significantly increasing the difficulty of detection. To address this challenge, this paper proposes a Transformer-based detection method for camouflaged objects to enhance the localization and recognition of camouflaged pests.The model introduces a Fine-Grained Score Predictor (FGSP) to guide queries toward potential foreground regions, combines it with a MaskMLP module to generate instance-aware pixel-level masks, and incorporates a Denoising Module and DropKey strategy to improve training stability and attention robustness, respectively.Experimental results show that the proposed model achieves AP scores of 36.31 and 75.07 on the COD10k and Locust datasets, outperforming Deformable DETR by 2.3% and 3.1%, respectively. Furthermore, on the Locust dataset, the proposed model improves the Recall by 6.15% and the F1-score by 6.52%. Ablation studies further validate the effectiveness of each component, demonstrating that the proposed method provides a robust solution for detecting camouflaged pests in complex field environments, and holds significant value for agricultural pest monitoring and crop protection.