The denoising process of fringe pattern can recover the boundary information of fringe pattern and thus improve the accuracy of fringe pattern 3D measurement results. In order to recover the boundary information of the fringe pattern as much as possible, a fringe pattern denoising method is proposed to improve the SwinIR neural network. First, the Inception module is introduced and the structure of the RSTB module in the network is optimized to improve the local feature extraction capability of the network. Second, multiple residual blocks are introduced to the overall structure of the network to alleviate the problem of gradient disappearance caused by over-deepening of the network. The de-noising performance was tested by using high-density area stripes. When the noise level is 50, the PSNR value of the improved SwinIR algorithm is 31.96, the SSIM value is 0.995 5, and the denoising time is 4.035 s. Moreover, the improved SwinIR algorithm is compared with seven other representative algorithms, and the results show that the denoising performance of this method is optimal at different noise levels.