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

Abstract: To address the challenges in semi-supervised target detection due to the scarcity of fire image labels and the complexity of backgrounds, a Distribution Adaptation and Dynamic Curriculum Pseudo-Labeling framework for Semi-supervised Fire Detection (DADCPL-SFD) was proposed，which consisted of four modules: teacher-student mutual learning framework, soft labeling, distribution adaptation and dynamic curriculum pseudo-labeling. 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. Then, soft labels were used to obtain more effective pseudo-label positive examples and optimize the semi-supervised learning process. Next, the distribution adaptation loss was introduced to minimize the data distribution difference between the source domain and the target domain, ensuring consistent model performance across different domains. Finally, a dynamic curriculum pseudo-label strategy, inspired by the concept of curriculum learning, was designed. The threshold was dynamically adjusted during different training periods based on the pseudo-label generation process to filter out more accurate pseudo-labels. Experimental results on the Dataset for Fire and Smoke detection (DFS) at various supervision ratios (1%, 2%, 5% and 10%) show that the mean Average Precision (mAP) of the proposed framework is improved by an average of 5.32 percentage points compared to the original supervised learning, and Average Precision at an Intersection over Union (IoU) threshold of 0.5 (AP50) is 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, soft labels, distribution adaptation；dynamic curriculum pseudo-labeling