| Summary: | The deep convolutional neural networks (CNNs) have been shown excellent performances for image denoising. However, the denoising CNN model trained with a specific noise level cannot deal with the images which have spatiotemporally variant random noise and low signal-to-noise ratio (SNR), such as seismic images. To this end, we propose a patch-based denoising CNN method, namely PDCNN. Specifically, we cluster the overlapping patches of noisy image into K classes where the image patches have close noise levels in each class, and then choose a suitable model for denoising the corresponding class from a series of well-trained CNN models. By embodying the structural statistics, we propose a CNN model selection criterion with a structural-dependent parameter. In contrast to the manual model selection process, the more accurate CNN model is chosen automatically and effectively. The capability of the PDCNN is demonstrated on synthetic and field seismic images. Experimental results show that the proposed method largely benefits from using multiple CNN models to jointly denoise, and leads to the satisfactory denoising performance in spatiotemporally variant seismic random noise reduction and structural signal preservation.
|
|---|
