<p>The fire detection method based on machine vision can predict the early fire situation of forest farm and reduce the workload of forest farm guards. In general, you can use a fixed view wireless image sensor to monitor the situation in the forest environment (within a certain time frame). The digital image segmentation method can be used to judge the flame component in the wireless image sensor image and realize the fire warning. However, the video captured by wireless image sensors is a multi-dimensional image that changes with time, so there will be some errors in flame segmentation, which will lead to error detection. In this paper, a flame detection method based on deep learning and Gaussian mixture model is designed based on the characteristics of flame motion and flame segmentation. The method can improve the reliability and correctness of flame detection by combining the result of flame segmentation with the binary image generated by dynamic detection. In this paper, a parallel scheme of binary image intersection operation and a series scheme of image segmentation and dynamic detection are proposed. In order to verify the reliability of the two schemes, experiments were carried out on the flame scene in the laboratory, the electric drive light source scene and the actual fire scene in the forest farm. Through experimental verification of flame videos obtained by wireless image sensors in different scenarios, the method described in this paper can distinguish flames in different scenarios and identify flames with certain risks, thereby further improving the fire recognition capability based on fixed view wireless image sensors.</p>