Fully convolutional one-stage object detection model for fire and smoke detection

Abstract

Fire and smoke detection is one of the most studied fields in the literature today. Considering the speed of climate change and its effects like fire starters, the field focuses on preventing potential disasters, saving the lives of people and wildlife, and protecting the environment by early detection systems and warnings. Especially early detection may have a vital effect to help slow down the spread of forest fires. Since the advent of deep learning, this task has been addressed into object detection methods, instead of sensor or detector based physical solutions. In this paper, we propose an anchor-free, lightweight, fully convolutional, one-stage object-detection-based deep neural network for fire and smoke detection. An anchor-free model eliminates the need for predefined, non-learnable anchor parameters. A lightweight model requires fewer computational resources, making it more environmentally friendly and better suited for such an ecological solution. It is also a one-stage model, trainable end-to-end without requiring a separate region proposal stage. Our proposed FCOS-based model integrates an EfficientNet-b3-based lightweight backbone to extract spatial information, a custom bidirectional Feature Pyramid Network built using Cross-Stage Partial (CSP) module of convolutional blocks and a Spatial Pyramid Pooling–Fast module, and a multi-scale Detection Head module built with depthwise-separable convolutional blocks and Convolutional Block Attention Modules. We discuss the background of object detection models, the differences and weaknesses of each method. We also propose a Faster R-CNN model as a two-stage object detector for an ablation study. The model is customized with MobileNet-v3 Large-based feature extractor network, an improved Region of Interest Head with a deeper convolutional structure and an improved Multi-layer Perceptron module for the Box Head detector. We used the Fire and Smoke Dataset from Roboflow for training and evaluation. Our FCOS-based primary model achieved an mAP@0.5 score of 58.5% and an F1 score of 60.3%, which are 3% lower than that of the previous state-of-the-art RT-DETR-based transformer network while using 20% fewer parameters and incurring 14% lower computational cost (in GFLOPs) compared to the base RT-DETR model; our model also achieves 2.7% lower mAP@0.5 score than previous state-of-the-art convolutional neural network based YOLOv8-m model while using 55% fewer parameters and 35% less computational cost in our benchmarks. We also achieved a mAP@0.5 score of 55.7% (2.8% lower than the FCOS-based model) with our custom Faster R-CNN model, which requires approximately the same computational cost as the FCOS-based model.