Efficient FCOS Model for Fire and Smoke Detection

Early detection of forest fires and industrial incidents is vital for minimizing damage to the environment and saving lives. In this paper, we propose a lightweight, anchorfree, one-stage FCOS-based object detection model for fire and smoke detection. Our model integrates an EfficientNet-b3based lightweight encoder, a custom bidirectional Feature Pyramid Network created with Cross-Stage Partial blocks and Spatial Pyramid Pooling-Fast (SPPF), and a multi-scale detection head built with depthwise-separable convolutions and convolutional attention mechanisms. We train and evaluate our model on the Fire and Smoke Detection Dataset from Roboflow. Our approach achieves an mAP@50 score of 58.5% and an F1 score of $60.3 \%$, only $3 \%$ lower than the previous state-of-the-art RT-DETR-based solution while using $20 \%$ fewer parameters.