A lightweight LBM detection transformer with multi-scale feature fusion for citrus-picking robots

To address false positives and missed detections caused by complex orchard backgrounds and small target regions during citrus harvesting, this study proposes a lightweight LBM Detection Transformer for citrus recognition in citrus-picking robots, integrating cross-scale feature fusion. Firstly, a large-kernel attention module (LKAM) is introduced for feature extraction, where multi-scale large-kernel convolutions are employed to comprehensively capture global, local, and texture information while maintaining the lightweight nature of the model. Secondly, a multi-scale linear attention (MSLA) Transformer encoder is incorporated to further enhance feature discrimination, thereby enabling the network to focus more accurately on citrus targets and effectively suppress background interference. Finally, a bidirectional feature pyramid neck (BFPN) is designed to achieve efficient cross-layer feature fusion by integrating deep and shallow representations, which compensates for target information loss caused by occlusion. Experimental results on citrus orchard scenes demonstrate that, compared with the baseline model, the proposed LBM Detection Transformer reduces the parameter count by 5.54 M while improving mAP by 4.69%, mmAP by 5.14%, and R by 3.66%. These results indicate that the proposed method provides an accurate and lightweight solution for intelligent citrus management and automated harvesting applications.