@article{Al-Wesabi2025, 
author = {Fahd N. Al-Wesabi and Abdulaziz Alhefdhi},
title = {Enhancing safety for blind and visually impaired people: intelligent fire detection using deep learning and the lemurs optimization algorithm},
year = {2025},
journal = {AIMS Mathematics},
volume = {10},
number = {9},
pages = {21617-21641},
keywords = {fire detection, lemurs optimization, image pre-processing, visually challenged people, artificial intelligence},
url = {https://www.sciopen.com/article/10.3934/math.2025961},
doi = {10.3934/math.2025961},
abstract = {An increased number of older people suffer from higher levels of cognitive and vision impairments, which usually results in loss of independence. Fire detection systems play a crucial role in providing timely alerts to blind and visually impaired (BVI) individuals during indoor emergencies. As fire detection is complex and critical for safety, deep learning (DL) has recently been adopted for precise recognition. Efficient algorithms are crucial for hardware-constrained devices like embedded systems, robots, and mobiles to ensure high performance with low power use. In this paper, an enhanced fire detection system for blind and visually challenged people using artificial intelligence (AI) and Lemurs Optimisation Algorithm (EFDBVCP-AILOA) model is proposed. The aim is to assist visually impaired individuals by using DL techniques. Primarily, the adaptive bilateral filtering (ABF) method is used to reduce noise while preserving essential edges in fire images. For feature extraction, the NASNetMobile method is employed to capture complex features from the image data. Furthermore, the EFDBVCP-AILOA method implements self‐attention with a convolutional neural network and long short-term memory (CNN-Sa-LSTM) model for classification. Finally, the Lemur's Optimisation (LO) model is employed as a parameter-tuning approach for the CNN-Sa-LSTM model. A wide-ranging experimentation of the EFDBVCP-AILOA approach is accomplished under the fire detection dataset. The comparative results of the EFDBVCP-AILOA approach demonstrated superior    a  c  c            u              y       of 97.07%, with    p  r  e            c              n      ,    r  e  c            a              l      , and              F              m      e      a      s      u      r      e       values of 96.98%, 97.07%, and 97.00%, respectively. The EFDBVCP-AILOA approach applied four performance metrics to evaluate and compared done with seven recent models.}
}