Abstract:

In order to address the safety hazards caused by low visual recognition accuracy of fire sources during coal mining, a fire source visual recognition method based on multi-feature fusion was proposed. Firstly, a complex monitoring system network topology was constructed to collect fire source images. Gaussian filtering and an improved frame difference method were used for image preprocessing to remove noise and capture dynamic changes in flame brightness, resulting in high-quality images. Secondly, in the preprocessed image, key features of the fire source were extracted, including color characteristics, motion trajectories, and shape contours. A multi-image feature fusion strategy was adopted, and Gaussian kernel functions and weighted summation mechanisms were used to fuse different features into more expressive and discriminative feature representations. Finally, by combining support vector machine classification with a region-based convolutional neural network object detection algorithm, accurate identification of fire sources could be achieved. The results show that the proposed method has a fire source recognition accuracy of over 87%, indicating a relatively high recognition accuracy. It can accurately capture fire source features in various complex scenarios and has high accuracy in identifying coal fires.

Key words: multi-feature fusion, visual recognition, fire source identification, fire source image, feature extraction, convolutional neural network