Dual-Guided Semi-Supervised Semantic Segmentation for Citrus Quality Evaluation.

Area of Science:

• Computer Vision
• Precision Agriculture
• Machine Learning

Background:

• Supervised semantic segmentation for citrus defect detection requires extensive labeled data, leading to high costs.
• Existing semi-supervised methods struggle with pseudo-label instability and poor feature discriminability for complex defects.

Purpose of the Study:

• To develop a dual-guided semi-supervised semantic segmentation model, UP-ETS, for efficient and accurate citrus surface defect detection.
• To address challenges in noise propagation and feature learning for complex citrus defect segmentation.

Main Methods:

• UP-ETS utilizes a Mean Teacher-Student framework incorporating Uncertainty Estimation (UE) with Kullback-Leibler divergence to stabilize pseudo-labels.
• Prototype Contrastive Learning (PCL) is employed to enhance feature discriminability for complex citrus surface textures and blurred boundaries.

Main Results:

• UP-ETS achieved improved Dice scores (87.76% vs. 85.57% baseline) with only 1/16 labeled data.
• The model demonstrated enhanced segmentation for difficult samples, including small targets and blurred regions.
• UP-ETS outperformed other semi-supervised models in performance, parameters, and inference speed, showing robust cross-dataset generalization.

Conclusions:

• The proposed UP-ETS model effectively mitigates the need for large annotated datasets in agricultural defect detection.
• UE and PCL synergistically create a more structured and discriminative feature space, proving their efficacy.
• UP-ETS offers significant practical value for agricultural deployment due to its efficiency and accuracy.