TL-FSD-YOLO11s: The Detection of Health Status of Passion Fruit using Transfer Learning with Improved YOLO11s

Yu Zhou^1*Sheng Xue¹Zhenye Li¹Chao Ni^1*Tingting Zhu¹Jian Wu²

• ¹Nanjing Forestry University, Nanjing, China
• ²Nantong Wealth Machinery Technical Co.,Ltd., Rugao, China

Accurately assessing the health condition of the passion fruit surface is crucial for maintaining market competitiveness. Traditional manual inspection is prone to subjective bias, often leading to inconsistent evaluations that reduce both market value and sales potential. Leveraging the capability of deep learning methods to automatically extract relevant features from regions of interest, this study proposes a novel neural network named Transferring Learning-FasterBlock-SCSA-MPDIoU-YOLO11s (TL-FSD-YOLO11s) for the real-time detection of passion fruit health conditions. Several key enhancements characterize the proposed network architecture. First, weights labeled "PFbest.pt" were pre-trained on a publicly available passion fruit dataset, significantly improving the model's feature extraction capability specific to passion fruit samples. Second, the original C3K2 module in YOLO11s was replaced with the FasterBlock C3K2 module to enhance detection speed. Third, a Spatial and Channel Squeeze-and-Excitation Attention (SCSA) module was integrated into the network's neck to improve the model's sensitivity in detecting subtle defects. Finally, the Complete Intersection over Union (CIoU) loss function in the original YOLO11s network was substituted with the Minimum Pixel Distance Intersection over Union (MPDIoU) loss function, which accelerates model convergence and enhances detection precision. Additionally, partial layers of the backbone were frozen to enhance the network's generalization capability and mitigate overfitting. Experimental results demonstrated that the TL-FSD-YOLO11s model substantially outperformed the baseline YOLOv11s model, achieving an improvement of 11.2% in precision, 8.2% in recall, 11.1% in mean Average Precision (mAP@0.5), and 10.6% in mAP@0.5-0.95. Moreover, by freezing selected backbone layers, the number of model parameters and weight size were reduced by 27.4% and 1.3 MB, respectively, compared to YOLOv11s. The inference speed of the TL-FSD-YOLO11s network was also improved by 1.9 ms. These findings confirm that the proposed TL-FSD-YOLO11s network can effectively perform real-time and precise assessment of passion fruit health conditions.