THE IMPROVED FFDNet METHOD FOR DENOISING FRINGE PATTERN

In the three-dimensional measurement technology of digital fringe projection, the existence of noise often leads to the loss of fringe edge information, reduces the accuracy of phase extraction, and ultimately affects the accuracy of measurement results. In order to better preserve the fringe edge information, an improved FFDNet (fast and flexible denoising convolutional neural network) neural network fringe pattern denoising method is proposed. The LeakyReLU activation function and DenseNet were used to optimize the network structure of FFDNet, thereby improving the effect of model regularization and the utilization of the network layer. The Experimental results show that compared with FFDNet, the denoising effect of improved FFDNet is increased by 1.87~2.61dB at different noise levels, and the number of parameters is reduced by 75%.