Performance Evaluations of the Deep Learning Models in Reference to Real-Time Fire and Smoke Detections Abilities

Abstract

This study is primarily focused on developing a real-time fire and smoke detection systems using two predominant deep learning models, viz. VGG16 and RESNET50. Throughout this work, both of these models are operated in parallel, processed the frames from the input video footage, and made all the essential predictions on each frame. In the course of the practical implementations of the models, they are trained and validated well using the Fire-Smoke-Dataset retrieved from the open source code DeepQuestAI. Over the course of 120 epochs, the model’s accuracy is found to be improved from 80% to approximately 96% during the training. Their generic systems are noticed to be able enough to classify the fire, smoke, and neutral cases in all the pre-recorded videos and real-time webcam footages. The results presented herewith are basically vowed to demonstrate the efficacy and workability of both of the models, and their working performances (RESNET50 outperforms VGG16, and achieves an accuracy range of 87.67% compared to 82%). In terms of their precision and recall aptitudes, and efficiency & effectivity in identifying the fire and smoke instances, both of the models are found to offer the satisfactory detections mechanisms. We believe that the present in-depth yet critical evaluations and assessments on the workability of the deep learning models can illuminate the potentialities and promising functions of the VGG16 and RESNET50 towards detecting the real-time fire and smoke which in turn are quite indispensable for promoting them in the public safety and security mechanisms installed in the wide range industrial and non-industrial delicate sectors.