In non-contact 3D optical measurement based on deformation fringe pattern analysis, phase distribution is extracted from the collected deformation fringe pattern to obtain the surface information of the measured shape. But the acquired fringe pattern contains noise in measurement, which affects the accuracy of extracting phase information. In order to remove the noise in fringe pattern better and faster, an improved U-net neural networks filtering algorithm based on deep learning is proposed. In the field of image denoising, U-net acquires few shallow features. The proposed method contains 1×1 parallel convolutional branches in the convolutional of U-net, which is used to obtain multi-scale feature. And adds 1,2, 3 1×1 parallel convolutional branches for experiment. Fringe pattern with high-density regions is used in the experiment, and the proposed method is compared with the state-of-the-art deep-learning fringe pattern denoising algorithm. The denoising effect of the proposed method is improved by 0.9%, the denoising efficiency is improved by 41.7% and the training time is reduced by 30.8%.