Pseudo-Label Refinement for Robust Wheat Head Segmentation via Two-Stage Hybrid Training

2025-12-06

AI Summary

To address the challenges of high pseudo-label noise and poor robustness to small objects and occlusions in wheat spike segmentation, this paper proposes a two-stage self-training framework based on dynamic pseudo-label refinement. Methodologically: (1) an adaptive confidence thresholding mechanism coupled with spatial consistency verification is introduced to suppress noise propagation; (2) a teacher–student iterative training paradigm is established under strong data augmentation—including multi-scale elastic deformation, color perturbation, and random occlusion; (3) the SegFormer-MiT-B4 architecture is adopted to enhance fine-grained feature representation. Evaluated on the Global Wheat Full Semantic Segmentation Competition, the method achieves a 4.2% improvement in mean Intersection-over-Union (mIoU) over the baseline, demonstrating significantly enhanced robustness and consistency in segmenting wheat spikes under complex field conditions—such as variable illumination, dense overlap, and minute spike structures.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

This extended abstract details our solution for the Global Wheat Full Semantic Segmentation Competition. We developed a systematic self-training framework. This framework combines a two-stage hybrid training strategy with extensive data augmentation. Our core model is SegFormer with a Mix Transformer (MiT-B4) backbone. We employ an iterative teacher-student loop. This loop progressively refines model accuracy. It also maximizes data utilization. Our method achieved competitive performance. This was evident on both the Development and Testing Phase datasets.

Problem

• Develops a self-training framework for wheat head segmentation
• Refines pseudo-labels via iterative teacher-student training loop
• Combines two-stage hybrid training with extensive data augmentation

Innovation

• Two-stage hybrid training strategy
• Iterative teacher-student loop refinement
• SegFormer with MiT-B4 backbone