LoCoNet: A Low-Complexity Convolutional Neural Network Model for Efficient Fire Detection in Outdoor Environments

Early Fire Detection (FD) is essential, yet preventing damage to human life and property presents challenges. This study introduces a reliable and fast FD framework using a new Convolutional Neural Network (CNN) model called Low-Complexity Network (LoCoNet). The LoCoNet model deals with color images of 24×24 pixels, highly decreasing memory usage and processing time. The structure of the LoCoNet model consists of three convolutional layers, each utilizing a kernel size of 1×1, followed by a max-pooling layer, effectively halving the data size. Next, a flattening layer transforms the data into a 1-D vector. Then, a fully connected dense layer follows, and a dropout layer randomly deactivates 50% of its neurons during training. Finally, the output layer classifies the images according to the probability of fires occurring, predicting whether there are fires. K-fold cross-validation with various K values divided the dataset into training and testing sets. Multiple CNN models were investigated, and their results were compared to estimate their performance. According to the experimental results, the proposed LoCoNet model surpasses others in accuracy, processing speed, and memory usage, achieving an accuracy of approximately 99%, consuming about 2.86 s in model training, and using only 81.25 KB of memory. Compared to related approaches, the proposed LoCoNet model significantly decreases computational complexity while achieving high accuracy with minimal processing time.