Abstract
Fire detection is crucial for safeguarding human life and property. To address the limitations of existing deep learning-based detectors-such as weak feature perception, information loss, high computational cost, and poor performance on small targets-this paper proposes an enhanced YOLOv7 model named CGDS-YOLO. The model introduces three key innovations: a CDP-ELAN module (fusing Coordinate Convolution, Diverse Branch Block, and Partial Convolution) for strengthened feature extraction, a Gathering-Distributing mechanism for improved multi-scale information fusion, and a SlimNeck structure to reduce parameters while retaining fine-grained details. Additionally, Normalized Wasserstein Distance is adopted to enhance small target detection. Experiments on a homemade smoke and flame dataset and the public Visdrone dataset show that CGDS-YOLO outperforms baseline models, improving mAP by 2.0% and 1.7%, respectively, while maintaining high computational efficiency.