Smoke Recognition based on Dictionary and BP Neural Network

In order to improve the video-based smoke detection rate in a low illumination environment, a segmentation and recognition method of smoke images based on the combination of the dictionary and backpropagation (BP) neural network is proposed. Firstly, an image enhancement model is established for low-frequency and high-frequency images obtained by wavelet decomposition on the input image. The neural network predicts the coefficients of low-frequency and high-frequency images in the enhancement model to improve the image.e. Typical smoke images are collected to construct the smoke data matrix, and then background subtraction is employed to extract the motion region. The candidate smoke area is segmented further by principal component analysis and smoke data dictionary matrix. The feature vector is constructed using the statistics of the gray level co-occurrence matrix after the wavelet decomposition of each layer image. Finally, smoke is recognized by BP neural network. The experimental results show that the segmentation of the smoke image is relatively more complete, and the recognition rate is higher.