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

Accurate identification of cherry maturity and precise detection of harvestable cherry contours are essential for the development of cherry-picking robots. However, occlusion, lighting variation, and blurriness in natural orchard environments present significant challenges for real-time semantic segmentation.

To address these issues, we propose a machine vision approach based on the PIDNet real-time semantic segmentation framework. Redundant loss functions and residual blocks were removed to improve efficiency, and SwiftFormer-XS was adopted as a lightweight backbone to reduce complexity and accelerate inference. A Swift Rep-parameterized Hybrid (SwiftRep-Hybrid) module was designed to integrate local convolutional features with global Transformer-based context, while a Light Fusion Enhance (LFE) module with bidirectional enhancement and bilinear interpolation was introduced to strengthen feature representation. Additionally, a post-processing module was employed to refine class determination and visualize maturity classification results.

The proposed model achieved a mean Intersection over Union (MIoU) of 72.2% and a pixel accuracy (PA) of 99.82%, surpassing state-of-the-art real-time segmentation models such as PIDNet, DDRNet, and Fast-SCNN. Furthermore, when deployed on an embedded Jetson TX2 platform, the model maintained competitive inference speed and accuracy, confirming its feasibility for real-world robotic harvesting applications.

This study presents a lightweight, accurate, and efficient solution for cherry maturity recognition and contour detection in robotic harvesting. The proposed approach enhances robustness under challenging agricultural conditions and shows strong potential for deployment in intelligent harvesting systems, contributing to the advancement of precision agriculture technologies.