Cherry-Net:Real-time segmentation algorithm of cherry maturity based on improved PIDNet

Jie Cui^1,2,3Lilian Zhang^1,2,3Lutao Gao^1,2,3Chunhui Bai^1,2,3Linnan Yang^1,2,3*

• ¹Yunnan Agricultural University, Kunming, China
• ²Yunnan Engineering Technology Research Center of Agricultural Big Data, Kunming, China
• ³Yunnan Engineering Research Center for Big Data Intelligent Information Processing of Green Agricultural Products, Kunming, China

Accurate identification of cherry maturity and precise detection of harvestable cherry contours are key technologies for cherry-picking robots. To address this challenge, this study proposes a machine vision approach based on the PIDNet real-time semantic segmentation model.The proposed method enables precise pixel-level segmentation and maturity classification, enhancing its applicability in practical harvesting scenarios. To improve model efficiency, redundant loss functions and residual blocks were removed. A lightweight SwiftFormer-XS was adopted as the backbone network, significantly reducing model complexity and improving inference speed. Furthermore, a Swift Rep-parameterized Hybrid(SwiftRep-Hybrid) module is introduced, combining local feature extraction from convolution with global context modeling from Transformers. This module enhances the model's ability to capture comprehensive feature information in complex agricultural environments. In addition, a Light Fusion Enhance(LFE) module is designed, incorporating a bidirectional enhancement mechanism and bilinear interpolation to reinforce image features. This design effectively addresses challenges such as occlusion, blurriness, and lighting variations. A post-processing module was also introduced to refine class determination and visualize maturity classification results, facilitating deployment on robotic harvesting platforms. Experimental results show that the proposed model achieved an MIoU of over 72.2% and a pixel accuracy(PA) of 99.82%, outperforming state-of-the-art real-time segmentation models such as PIDNet, DDRNet, and Fast-SCNN. While maintaining high inference speed and low model complexity.The proposed approach provides a lightweight, accurate, and efficient solution for real-time semantic segmentation in cherry-picking robotics. It demonstrates strong potential for deployment in intelligent harvesting systems and contributes to the advancement of precision agriculture technologies.