Smoke Detection Algorithm Based on Improved Feature Fusion

Abstract

The smoke detection algorithm has wide application value in fire warning, environmental monitoring, and other fields. To improve the accuracy and real-time performance of smoke detection, a smoke detection algorithm based on improved feature fusion is proposed. A new feature matrix is constructed by combining the Haar-like algorithm and the LBP algorithm. To improve the two feature extraction methods, the study compares the pixel values of the neighborhood and the central pixel and weights the pixels of each region according to the difference. The improved LBP matrix is obtained. The longitudinal, transverse, diagonal, and cross-shaped feature differences of the Haar template are extracted, and these feature differences are used as the weighting coefficients of the Haar feature matrix. Then, the extracted features are input into the improved YOLOv7, and the efficient pyramid split attention network (EPSANet) is introduced into YOLOv7. Meanwhile, the dual compressed attention mechanism (shuffle attention network, SANet) integrates spatial channels. EPSANet extracts multilevel smoke feature information by squeezing and allocating attention to features of different scales through a pyramid-shaped structure. SANet further enhances the model’s focus on key features and reduces background interference by integrating a dual compression attention mechanism of space and channels. The results show that the detection accuracy of the research design method in different practical application scenarios is above 85%, and the first detection time is also within 5s. The detection error value of the research method mainly fluctuates in the range of 0.02∼0.04, which has high detection accuracy and detection efficiency. In summary, it can be seen that the smoke detection method designed in this study can provide more reliable technical support for fire warning, environmental monitoring, and other fields.