基于对比学习和伪异常合成的无监督火灾检测

传统的火灾检测方法大多基于目标检测技术, 存在火灾样本获取难度高、人工标注成本高的问题. 为解决该问题, 本研究提出了一种基于对比学习和伪异常合成的无监督火灾检测模型. 为了实现无监督图像特征学习, 提出了交叉输入对比学习模块. 然后, 引入了一个记忆原型学习正常场景图像的特征分布, 通过特征重建实现对火灾场景的判别. 并且, 提出了伪异常火灾场景合成方法和基于欧氏距离的异常特征区分损失, 使模型对于火灾场景具有针对性. 根据实验表明, 我们的方法在Fire-Flame-Dataset和Fire-Detection-Image-Dataset两个公开火灾检测数据集上的图像级AUC分别达到89.86%和89.56%, 优于PatchCore、PANDA、Mean-Shift等主流图像异常检测算法.

Unsupervised Fire Detection Based on Contrastive Learning and Synthetic Pseudo Anomalies

Traditional fire detection methods are mostly based on object detection techniques, which suffer from difficulties in acquiring fire samples and high manual annotation costs. To address this issue, this study proposes an unsupervised fire detection model based on contrastive learning and synthetic pseudo anomalies. A cross-input contrastive learning module is proposed for achieving unsupervised image feature learning. Then, a memory prototype that learns the feature distribution of normal scene images to discriminate fire scenes through feature reconstruction is introduced. Moreover, a method for synthesizing pseudo anomaly fire scenes and an anomaly feature discrimination loss based on Euclidean distance are proposed, making the model more targeted toward fire scenes. Experimental results demonstrate that the proposed method achieves an image-level AUC of 89.86% and 89.56% on the publicly available Fire-Flame-Dataset and Fire-Detection-Image-Dataset, respectively, surpassing mainstream image anomaly detection algorithms such as PatchCore, PANDA, and Mean-Shift.