Abstract To achieve rapid and high-accuracy recognition of peach maturity in complex orchard environments characterized by leaf occlusions, noise interference, and fruit overlapping, an enhanced YOLOv10n-based maturity detection method named YOLOv10n-MSI is proposed. First, a Median-Enhanced Channel Attention Module (MECAS), which integrates median pooling and multi-scale convolution, is incorporated into the feature extraction network to enhance the ability to extract features of occluded peaches. Then, a Switchable Atrous Convolution (SAConv) is adopted in the feature downsampling layer to improve the efficiency of spatial feature fusion. Furthermore, the Inner-IoU loss function is employed to optimize localization performance through an auxiliary bounding box mechanism. Experimental results demonstrate that the YOLOv10n-MSI model achieves a detection precision of 93.5%, a recall of 93.5%, and a mean average precision (mAP@0.5) of 97.7%, with a computational cost of 6.3 GFLOPs and 2.6 M parameters. Compared to the baseline YOLOv10n, the proposed model improves precision by 1.7 percentage points, recall by 4.6 percentage points, and mAP@0.5 by 1.6 percentage points. In comparisons with models such as YOLOv8s, YOLOv9s, and SSD, YOLOv10n-MSI reduces the number of parameters by 31.1% and computational cost by 2.4 GFLOPs (a 27.6% reduction compared to YOLOv8s). The proposed model balances accuracy and speed in peach maturity detection, demonstrating superior overall performance and providing technical support for intelligent real-time peach harvesting.
Yutao et al. (Fri,) studied this question.