YOLOv10-kiwi: a YOLOv10-based lightweight kiwifruit detection model in trellised orchards

To address the challenge of real-time kiwifruit detection in trellised orchards, this paper proposes YOLOv10-Kiwi, a lightweight detection model optimized for resource-constrained devices. First, a more compact network is developed by adjusting the scaling factors of the YOLOv10n architecture. Second, to further reduce model complexity, a novel C2fDualHet module is proposed by integrating two consecutive Heterogeneous Kernel Convolution (HetConv) layers as a replacement for the traditional Bottleneck structure. This replacement enables parallel processing and enhances feature extraction efficiency. By combining heterogeneous kernels in sequence, C2fDualHet captures both local and global features while significantly lowering parameter count and computational cost. To mitigate potential accuracy loss due to lightweighting, a Cross-Channel Fusion Module (CCFM) is introduced in the neck network. This module incorporates four additional convolutional layers to adjust channel dimensions and strengthen cross-channel information flow, thereby enhancing multi-scale feature integration. In addition, a MPDIoU loss function is introduced to overcome the limitations of the traditional CIoU in terms of aspect ratio mismatch and bounding box regression, accelerating convergence and improving detection accuracy. Experimental results demonstrate that YOLOv10-Kiwi achieves a model size of only 2.02 MB, with 0.51M parameters and 2.1 GFLOPs, representing reductions of 80.34%, 81.11%, and 68.18%, respectively, compared to the YOLOv10n baseline. On a self-constructed kiwifruit dataset, the model achieves 93.6% mAP@50 and an inference speed of 74 FPS. ​​YOLOv10-Kiwi offers an efficient solution for automated kiwifruit detection on low-power agricultural robots.