Towards Safer Environments: CNN-LSTM Framework for Accurate Fire Detection

Fire detection systems are highly important in giving early warnings and emergency responses. However, this sector often encounters a number of challenges due to a good level of false alarms that cause unnecessary mobilization of resources and a lack of belief in the system. This paper is meant for the presentation of a strong framework in the reduction of fire detection false alarms using an advanced classification model. The proposed system contains three crucial phases: preprocessing, feature extraction, and classification. It uses state-of-the-art image enhancement, noise reduction, and background subtraction techniques with adaptive thresholding during the system's preprocessing phase of the input data to enhance its quality. In this way, it ensures that features irrelevant to fire detection are eliminated and that the quality of data about fire either captured by surveillance footage or sensor readings is improved. Deep learning techniques are proposed in feature extraction, specifically Convolutional Neural Networks that automatically capture intricate patterns associated with fire occurrences, flame color, texture, and motion dynamics. This classification uses the architecture of the LSTM to concentrate on critical fire-related features, which will further enhance the accuracy in distinguishing events between fire and non-fire conditions. Training and testing datasets are a rich set of fire and non-fire instances from publicly available fire surveillance datasets, enriched with real-world smoke sensor and thermal camera data. Standard metrics of performance like accuracy, precision, recall, F1-score, and false alarm rate will be considered in order to validate the proposed framework comprehensively, covering fire detection and its false alarm reduction. The experimental results showed that the false positives drastically decreased, proving that the framework indeed supported enhancing the reliability of fire detection systems.