首页|Structured Deep Image Prior with Interscale Thresholding
Structured Deep Image Prior with Interscale Thresholding
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NSTL
This work proposes a novel image denoising technique inspired by the deep image prior (DIP) method. Our contribution is to increase the interpretability of the network by proposing to use the Stein's unbiased risk estimator (SURE) to realize self-supervised learning of image restoration. As a result, the number of parameters is decreased while keeping the performance. The conventional DIP accepts random input to generate restored image and has an advantage that no training data is requested. However, there is a problem that the interpretability is low. As a result, the network should prepare redundant design parameters. In this work, we replace the loss function from mean-squared error (MSE) to SURE. This replacement allows us to interpret the reason why a random input is needed. As well, we also introduce the interscale linear expansion of the thresholding (LET) in our network to exploit the extracted features. In order to avoid using group delay compensation, we construct a structured DIP by using hierarchical non-separable oversampled lapped transform (NSOLT) with the symmetric property. By showing some simulation results, the significance of the proposed method is verified.
NSTL
