Detection of Anomalous Fire using Deep Learning Techniques

One of the most damaging anomalous occurrences is fire. Both human lives and property are severely damaged and destroyed by it. Deep Learning has recently demonstrated promising outcomes in a number of classification and detection research projects. A better technique is still needed for effective fire and smoke detection, though. It is largely because of its ambiguous shape, texture, and potential for appearing in several shapes in day-to-day existence. In order to identify smoke and fire in films or photos, a novel, lightweight, real-time convolutional neural network is proposed in this study. Exiting datasets are constrained or artificially produced for testing.The validation process for this study's challenging planned dataset, which includes the majority of fire and smoke event scenarios, has been completed. An outcome that incorporates bounding box localization of the fire and smoke regions has been obtained on a highly diversified as well as newly introduced and targeted early fire detection image dataset. In comparison to RetinaNet, MobileNet, InceptionNet, and FireNet on the available dataset, the suggested approaches with modified EfficientDet obtained roughly 80% in terms of Average Precision (AP). Future research in this field may benefit from the proposed dataset, and implementing the evaluation by the suggested method to a real-world scenario would be beneficial.