A lightweight detection method for citrus fruits based on improved YOLO11 in dense occlusion environments

Abstract

To address the challenges of low detection accuracy, high miss rates, and limited model lightweightness arising from dense fruit distribution, foliage occlusion, and small fruit size during citrus fruit localisation and recognition in complex orchard environments, this study proposes a lightweight small-object citrus fruit detection model based on an improved YOLO11 architecture, termed SOCD-YOLO (YOLO for small object citrus detection). Firstly, a DLTBlock is designed by integrating element-wise multiplication with a triplet attention mechanism to reconstruct the C3k2 module, thereby enhancing the nonlinear fusion of high-dimensional features. This design effectively suppresses interference from occluding foliage and complex backgrounds, and improving the robustness of fruit target recognition. Secondly, the traditional downsampling operation is replaced with the adaptive downsampling block module, which effectively alleviates information loss during feature propagation for small-object features, while simultaneously reducing model parameters and computing complexity, thus enhancing small-object recognition accuracy. Finally, a lightweight P2-focused pyramid structure is constructed to further enhance the model’s capability in identifying small and densely distributed objects, while significantly reducing the missed detection rate. Experimental findings indicate that, on the CitDet dataset, the proposed model achieves improvements of 4.3%, 6.3%, and 5.6% in Precision, Recall, and mean average precision, respectively, while reducing the parameter count and model size to 1.5 M and 3.5 MB. Moreover, the FPS reaches 103.5 frames/s. On the Tomato and PASCAL VOC 2007 datasets, overall performance is consistently improved. Compared to existing object detection models, SOCD-YOLO demonstrates enhanced performance in terms of citrus fruit detection accuracy and robustness, providing a valuable reference for artificial intelligence based real-time fruit detection and position measurement in densely occluded environments.