基于变分推理半监督学习的玉米雄穗实例分割方法

Instance Segmentation of Corn Tassels Based on Adaptive Variational Bayesian Semi-supervised Learning

摘要:

针对玉米雄穗实例分割任务中标注数据制作耗时费力、易受主观因素影响,以及田间复杂场景导致预测不确定性高的问题,本文提出一种自适应变分推理贝叶斯实例分割模型（AVB-IS）。AVB-IS将贝叶斯概率框架与深度学习相融合,通过变分推理学习玉米雄穗特征分布,并通过自适应先验优化潜在空间,使模型在单次前向传播中即可预测不确定性。AVB-IS利用不确定性信息增强特征表达,提升伪标签质量和模型鲁棒性。结果表明,AVB-IS在多种骨干网络中实现性能与效率的良好平衡,平均精度（AP）达到49.35%,实例分割能力优于当前主流模型;同时,通过调节KL散度权重,模型展现出更快且更稳定的收敛特性。基于AVB-IS模型半监督学习（AVB-IS-SSL）框架下,仅使用50%的标注数据可获得接近全监督的效果,显著降低了对标注数据的依赖。本研究为玉米雄穗智能检测提供了可靠技术方案。

Abstract:

Aiming to address the problems of time-consuming and labor-intensive annotation in corn tassel instance segmentation,susceptibility to subjective factors,and high prediction uncertainty caused by complex field scenes,an adaptive variational-inference Bayesian instance segmentation model (AVB-IS) was proposed. AVB-IS integrated Bayesian probabilistic analysis with deep learning,learned the feature distribution of corn tassels through variational inference,and optimized the latent space with adaptive priors,enabling the model to predict uncertainty in a single forward pass. AVB-IS leveraged uncertainty information to enhance feature representation,improving pseudo-label quality and model robustness. Experimental results demonstrated that AVB-IS achieved an effective balance between performance and efficiency across diverse backbone networks,attaining an average precision (AP) of 49.35% and delivering instance segmentation performance that surpassed mainstream models. Moreover,adjusting the KL divergence weight enabled faster and more stable convergence. Under the framework of semi-supervised learning with the AVB-IS model（AVB-IS-SSL）,the model achieved performance comparable to full supervision using only 50% of the annotated data,substantially reducing dependence on manual labeling. The research result can offer a robust technical solution for intelligent maize tassel detection and hold significant value for advancing agricultural intelligence.