Zhao, Ziang ![]() ![]() Item availability restricted. |
Preview |
PDF (Thesis)
- Accepted Post-Print Version
Download (16MB) | Preview |
![]() |
PDF (Cardiff University Electronic Publication Form)
- Supplemental Material
Restricted to Repository staff only Download (196kB) |
Abstract
Fruits play a fundamental role in human nutrition, serving as a key source of essential vitamins and minerals. However, the global fruit industry is facing a significant challenge: the shortage of labour for harvesting, which remains a predominantly manual task. Automated fruit-harvesting robots present a promising solution to address this labour gap and maintain stable production. These robots can operate continuously without fatigue, yet they struggle to accurately assess fruit ripeness, which is a critical factor influencing harvest quality and timing. While numerous laboratory-based techniques for evaluating ripeness have been developed, their application in field settings is limited due to the complex and variable conditions of real-world orchards. To address these challenges, this thesis explores deep learning for determining fruit ripeness through vision models and high-quality fruit image datasets. Specifically, this thesis introduces NinePeach, a large dataset of peach images, and PeachSOLO, a one-stage model designed for peach instance segmentation. PeachSOLO achieves an average precision (AP) of 72.12, surpassing Mask R-CNN (69.91 AP). This thesis then proposes LightStraw, a lightweight model for strawberry instance segmentation. It requires considerably fewer parameters (17.42M) and floating-point operations (78.3G) than Mask R-CNN (35.08M/877.4G). This thesis also combines peach and strawberry images into a single dataset and proposes a query-based segmentation model FruitQuery. FruitQuery achieves the best AP of 67.02 with only 14.08M parameters, outperforming 13 other models with 33 variants, including three series of YOLO. Finally, this thesis develops AppleSSL, a self-supervised method for assessing in-field apple ripeness under occlusion. Using less than 1% labelled images, AppleSSL reconstructs obscured parts and provides ripeness scores from 0.0 to 1.0, surpassing 15 other self-supervised methods and one supervised method. Overall, this thesis demonstrates that deep learning can enable practical, accurate, and efficient ripeness estimation in real-world environments, supporting robotic fruit picking and contributing to smart, precision agriculture.
Item Type: | Thesis (PhD) |
---|---|
Status: | Unpublished |
Schools: | Schools > Engineering |
Uncontrolled Keywords: | 1. fruit 2. ripeness 3. deep learning 4. segmentation |
Date of First Compliant Deposit: | 29 September 2025 |
Last Modified: | 29 Sep 2025 12:57 |
URI: | https://orca.cardiff.ac.uk/id/eprint/181374 |
Actions (repository staff only)
![]() |
Edit Item |