Abstract:

To address the challenges in semi-supervised object detection due to the lack of fire image labels and the complexity and diversity of background， a Distribution Adaptation and Dynamic Curriculum Pseudo-Label framework for Semi-supervised Fire Detection （DADCPL-SFD） was proposed， which consisted of four parts： teacher-student Mutual Learning （ML） framework， Soft Label （SL）， distribution adaptation and dynamic curriculum pseudo-label. Firstly， the semi-supervised learning paradigm of teacher-student mutual learning framework was adopted to replace the fully supervised learning paradigm of YOLOv5-l for the scenario with few data labels. Secondly， soft labels were used to obtain more effective pseudo-label positive examples and optimize the semi-supervised learning process. Thirdly， the distribution adaptation loss was introduced to reduce data distribution difference between the source domain and the target domain， thereby ensuring consistent model performance across different domains. Finally， a dynamic curriculum pseudo-label strategy， inspired by the concept of curriculum learning， was designed to adjust the threshold according to the pseudo-label generation condition dynamically during different training periods， so as to filter more reasonable pseudo-labels. Experimental results on the Dataset for Fire and Smoke detection （DFS） at various supervision ratios （1%， 2%， 5%， and 10%） show that compared to the supervised learning， the proposed framework has the mean Average Precision （mAP） improved by an average of 5.32 percentage points， and the Average Precision （AP） at an Intersection over Union （IoU） threshold of 0.5 improved by an average of 11.87 percentage points， which fully demonstrates the efficiency and accuracy of DADCPL-SFD.

Key words: fire detection, semi-supervised object detection, teacher-student Mutual Learning (ML), Soft Label (SL), distribution adaptation, dynamic curriculum pseudo-label

摘要：

针对火灾图像标签过少、背景复杂多样导致的半监督目标检测困难的问题，提出一种用于半监督火灾检测的分布自适应和动态课程伪标签框架（DADCPL-SFD）。该框架主要由师生互学（ML）框架、软标签（SL）、分布自适应和动态课程伪标签这4个部分组成。首先，采用师生互学框架的半监督学习范式替换YOLOv5-l的全监督学习范式，以应对数据标签少的场景；其次，采用软标签以获取更多有效的伪标签正例，优化半监督学习过程；再次，引入分布自适应损失，减小源域和目标域的数据分布差异，使模型在不同域上表现一致；最后，设计一种基于课程思想的动态课程伪标签策略，根据伪标签生成的情况在不同训练时期调整阈值，以筛选更合理的伪标签。在火焰和烟雾数据集（DFS）多个监督比例上（1%、2%、5%和10%）的实验结果表明，相较于全监督学习，所提框架的平均精度均值（mAP）平均提升了5.32个百分点，在交并比（IoU）阈值为0.5下的平均精度（AP）平均提升了11.87个百分点，充分验证了DADCPL-SFD的高效性和准确性。

关键词: 火灾检测, 半监督目标检测, 师生互学, 软标签, 分布自适应, 动态课程伪标签