Abstract
The developing elderly population undergoes a high level of eyesight and mental impairment, which frequently results in a defeat of independence. That kind of person should do vital daily tasks like heating and cooking, with methods and devices intended for visually unaffected persons, which does not consider the requirements of people with blind and intellectual impairment. Innovative technology needs the proper techniques for perceiving fires as rapidly as possible to avert damages. Initial fire recognition and notification models deliver fire inhibition and protection information to visually challenged individuals in an emergency if a fire happens indoors. Using an early fire recognition and warning model for blind individuals can decrease the number of victims, the number of losses, and, most essentially, early deaths. Recently, the growth of the fire recognition approach using artificial intelligence (AI) has advanced in helping blind people. This manuscript presents a Smart Fire Detection System for Assisting the Blind Using Attention Mechanism-Driven Recurrent Neural Network and Seahorse Optimizer Algorithm (SFDAB-ARNNSHO). The main intention of the SFDAB-ARNNSHO method is to detect and classify fire for blind people. To achieve this, the proposed SFDAB-ARNNSHO model performs image pre-processing by utilizing the sobel filtering (SF) model to remove noise in input data. Furthermore, the fusion of feature extraction comprises three methods, EfficientNetB7, CapsNet, and ShuffleNetV2. Furthermore, the SFDAB-ARNNSHO model performs fire detection and classification using stacked two-layer bidirectional long short-term memory with attention mechanism (SBiLSTM-AM) technique. Finally, the parameter tuning of the SBiLSTM-AM method is accomplished by implementing the seahorse optimizer (SHO) technique. The simulation validation of the SFDAB-ARNNSHO methodology is examined under the fire detection dataset, and the outcomes are measured using various measures. The performance validation of the SFDAB-ARNNSHO methodology portrayed a superior accuracy value of 99.30% over existing models under diverse measures.