Muzzle flame is one of the bases to study the performance of propellants, the most common method to study muzzle flame is through image processing technology, the geometric characteristic parameters of muzzle flame can be accurately measured by using high-speed camera method in the process of projectile firing, and then the performance of propellants can be measured. In the high-speed camera method,this paper abandons the traditional image segmentation algorithm and uses semantic segmentation model based on UNET to segment muzzle flame in order to realize the separation of muzzle flame and complex background environment. A deeply separable convolution and residual structure are introduced, and the Unet semantic segmentation model is optimized in order to improve the segmentation effect of the muzzle flame. After comparison and experiment verification, the accuracy of the model has increased from 0.965 to 0.979.