Abstract:

The S-FCN fire image detection method based on feature engineering was proposed to address the issues of high computational complexity and poor real-time performance of deep learning algorithms for fire image detection in complex backgrounds. Firstly, this method extracted color features from images in multiple color spaces and reduced the dimensionality of these features using mutual information. Secondly, the network structure of the deep learning model was simplified by using a single hidden layer of a fully connected network as its backbone. The color features in multiple color spaces can better represent fire smoke and flames, and reducing the dimensionality of color features in multiple color spaces effectively reduces the redundancy of input features. The single hidden layer fully connected network can significantly reduce the number of parameters during the model propagation process. Finally, this method was evaluated on a real and complex background fire image dataset. The experimental results show that the detection accuracy achieved by this method is 93.83%, and the real-time frame rate is 10 869 f/s. This method achieves high accuracy and high-speed fire image detection in complex scenes.

Key words: feature engineering, single hidden layer fully connected network(S-FCN), fire image, detection method, color space, feature dimensionality reduction