dc.description.abstract |
Destructive elements like fire quickly affect us with detrimental effects and to prevent their devastating damages, some preventive measures need to be taken immediately. As a large number of live are claimed due to fire-caused accidents, being able to detect fires in earlier stages might increase the chance of survival in any emergencies. Computer vision is proven to be beneficial in terms of almost accurate scenario-based classification tasks and can also be used for fire-scene detections. There remains a concern of scenario-based detections that some fire-like objects might also get classified as fire due to lack of proper training of the Deep Learning models. To mitigate this situation, an enhanced dataset is prepared after manual observation and careful inspection to distinguish between real and fake fire images. Among three different tested Transfer Learning models, Xception is proven to achieve the highest accuracy with a score of 97.24% over the custom augmented dataset consisting 4000 images of three classes: Fire, Fire-Like and Non-Fire. Analyzing the classification accuracy of Xception, this study suggests using the proposed method for differentiating between real fire and fire-like image scenarios. © 2023 IEEE. |
en_US |