GP-DETR: a lightweight real-time intelligent model for near-background colour pepper detection in complex agricultural environments

Green pepper detection faces substantial challenges due to near-color background interference, target overlap occlusion, and heterogeneous target scale distributions in greenhouse environments. Therefore, this study proposes a novel lightweight green pepper detection algorithm called Green Pepper Detection Transformer (GP-DETR) based on an improved Real-Time Detection Transformer (RT-DETR). It specifically aims to address the difficulties of detecting occluded and scaled targets in near-color backgrounds. First, the Generalized Efficient Layer Aggregation Network (GELAN) serves as the feature extraction backbone, integrating the Adaptive Downsampling Operator (ADown) to optimize feature transfer efficiency, along with the Efficient Local Attention (ELA) mechanism embedded within specific feature extraction layer convolution modules. Second, an improved Multi-scale ZhuYi Mechanism Attention-based Intrascale Feature Interaction (MsZZMAIFI) module is proposed, enhancing the model’s detection performance for multi-scale targets. Third, a novel Bilateral Enhancement Feature Pyramid Network (BiE-FPN) is designed based on the Bilateral Enhancement (BiE) module, which enhances the network’s discrimination ability for near-color targets. Last, the Re -parameterized Cross-Stage Partial Block with Bottleneck (RCSPBN) module with re-parameterization characteristics is proposed, which balances model performance and computational efficiency through a multi-branch structure in the training phase and a structure-simplified parameterization in the inference phase, thereby meeting the lightweight deployment requirements. The experimental results demonstrate that GP-DETR achieves a [email protected] of 96.4 % with an inference speed of 100 frames per second (FPS). The parameters, computational cost, and model Size are 43.4 M, 36.2 GFLOPs, and 23.3 MB respectively, which are reduced by 43 %, 37 %, and 40 % compared to the original RT-DETR. Comparative experiments demonstrate that it has advantages over similar mainstream algorithms in the YOLO series. The improved model demonstrates enhanced efficiency and accuracy in completing multi-scale feature detection in complex near-color backgrounds while meeting lightweight real-time requirements. This research provides valuable insights and novel approaches for solving target recognition problems in agricultural near-color backgrounds.