A deep learning model for real-time recognition of immature persimmons in complex field scenarios

Immature persimmons in natural field environments exhibit characteristics such as weak color distinctiveness, small target size, and frequent occlusion by branches and leaves, as well as interference caused by illumination variations. These factors significantly reduce the accuracy of real time detection.

To address this issue, this study proposes a real-time detection method for immature persimmons based on an improved YOLOv12n framework. Specifically, an image dataset of immature persimmon fruits was constructed under complex real-field conditions, encompassing diverse lighting conditions, occlusion levels, and shooting viewpoints. Multiple data augmentation strategies were employed to enhance sample diversity and improve the generalization capability of the model. Furthermore, MobileViTv3, CBAM, and DySample modules were integrated into the YOLOv12n architecture to simultaneously improve detection accuracy and inference speed. The effectiveness of MCD-YOLOv12n was comprehensively validated through Grad-CAM visualization, ablation studies, and comparative experiments, and its performance was evaluated against various classical models across different real-world scenarios.

Experimental results demonstrate that MCD-YOLOv12n achieves a Precision of 96.30%, a Recall of 89.90%, and an mAP of 95.30% on the persimmon fruit dataset, while maintaining a real-time inference speed of 72.5FPS. Moreover, the proposed model significantly outperforms mainstream object detection approaches, including DETR, EfficientDet, Faster R-CNN, YOLOv10n, and YOLOv12n, thereby satisfying the practical requirements of field operations.

These findings contribute to the advancement of smart agriculture and indicate strong application potential for real-time immature persimmon fruit detection and early-stage precision field management.