The dense smoke generated during the launching process of large caliber artillery can significantly degrade the imaging quality of high-speed camera systems, thereby affecting the accuracy of motion parameter measurement based on image analysis. To tackle this critical issue, this paper applies CycleGAN to the task of smoke removal in artillery launching images, in order to achieve end-to-end image enhancement with unpaired training images. Compared with traditional physics model-based smoke removal methods, this method autonomously learns the nonlinear transmission map between smoke distribution and clear images through adversarial training. Through subjective visual evaluation and quantitative analysis of objective metrics (Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM)), this method outperforms traditional enhancement algorithms such as Gamma correction and dark channel prior in detail restoration, with an average improvement of 3.48 dB in PSNR and 1.1% in SSIM.