PGLD-YOLO: a lightweight algorithm for pomegranate fruit localisation and recognition

Accurate localisation and recognition of pomegranate fruits in images with background interference are crucial for improving the efficiency of automated harvesting. To address the issues of excessive model parameters, high computational complexity, and inadequate detection accuracy of the existing pomegranate fruit detection algorithms, this study proposes a lightweight pomegranate fruit detection algorithm, You Only Look Once (YOLO) for Pomegranate Lightweight Detection (PGLD-YOLO), based on an enhanced YOLOv10s framework. First, to reduce the model’s size, parameter count, and computational complexity, the lightweight ShuffleNetV2 network is employed to reconstruct the YOLOv10s backbone, thereby substantially reducing the memory usage and computational cost while simultaneously enhancing the feature extraction. Second, to mitigate the impact of occlusion factors in the background and strengthen multi-scale feature fusion, the C2f_LEMA module is introduced into the neck network, combining partial convolution with an efficient multi-scale attention mechanism. This enhancement improves the model’s focus on the target regions, increases detection accuracy and localisation precision, and further bolsters the model’s robustness to some extent. Finally, to further reduce the model’s parameter count and size, the GroupNorm and Shared Head (GNSH) detection head is designed, incorporating shared convolutional layers and a fusion group normalisation strategy, thus effectively achieving architectural overhead. The experiment results demonstrate that the improved model achieves a mean average precision of 92.6% on the Pomegranate Images dataset, while the parameter count and computational complexity are reduced to 4.7M and 13.8G, respectively, resulting in a model size of 9.9 MB. The generalisation capability was simultaneously validated on the Apple Object Detection and PASCAL VOC 2007 datasets. Compared with other mainstream detection algorithms, it achieves a superior balance between detection accuracy, localisation precision, and model complexity, providing a robust and lightweight reference for pomegranate fruit.