The invention discloses a smoke image segmentation and recognition method based on a dictionary and a BP neural network. The method comprises the steps that 1, an image processor carries out enhancement processing on an input image; 2, the image processor compares the enhanced image with a smoke dictionary matrix to calculate a suspected smoke area; 3, the image processor performs multi-scale wavelet decomposition on the obtained suspected smoke image; and the statistics of the gray level co-occurrence matrix of each layer of image obtained after multi-scale wavelet decomposition are calculated. A texture feature vector is constructed by adopting the calculated statistics of the gray level co-occurrence matrix of each layer of image. The texture feature vector is inputted into a smoke identification BP neural network to obtain a smoke identification result. The method is simple in step, convenient to implement, high in smoke recognition efficiency and high in accuracy.