Fire and Smoke Detection Based on Improved YOLOV11

Fire and smoke detection is an important measure to ensure the safety of people’s lives and property, as well as a crucial link in maintaining ecological balance and supporting scientific research. Traditional object detection methods rely more on manually designed features and rules. Although they are relatively simple to implement, their performance is limited in complex and variable practical applications. In contrast, deep learning-based methods can automatically learn deep features in data and have higher accuracy and stronger generalization ability. However, complex backgrounds, large environmental changes, and data requirements pose great challenges to high-precision outdoor smoke detection. To address these issues, this paper proposes an improved model, YOLOV11-DH3, based on YOLOV11. In this paper, the core DCN2 (Deformable Convolutional Networks2) of the YOLOV11 Head is replaced with the DCN3 module to form a new detection head. In addition, the loss function CIOU in YOLOV11 is replaced with IOU to consider the irregular shape of fire and smoke and the problem of multi-scale targets. To evaluate the performance of the algorithm, comprehensive experiments were conducted on two distinct datasets: a public fire and smoke dataset provided by Baidu Paddle featuring close-range views and a public wildfire smoke dataset from the YOLO official website with distant outdoor perspectives. The experimental results show that on the Baidu Paddle dataset, the average accuracy of the model is improved by 1.4% compared to the original model, reaching 58%, the F1 score is improved by 2%, reaching 58%, with a precision of 91.6% and recall of 90%. Our cross-dataset analysis provides valuable insights into model performance across different detection scenarios. The proposed model demonstrates the ability to accurately detect fire and smoke in complex backgrounds, and this progress is of great significance for protecting people’s lives and maintaining ecological balance.