A Seismic Image Denoising Method Based on Kernel-prediction CNN Architecture

To filter noises and preserve the details of seismic images, a denoising method based on kernel prediction convolution neural network (CNN) architecture is proposed. The method consists of two convolution layers and a residual connection, containing a source sensing encoder, a spatial feature extractor and a kernel predictor. The scalar kernel was normalized by the softmax function to obtain the denoised images. In addition, to avoid excessive blur at the expense of image details, the authors put forward the concept of asymmetric loss function, which would enable users to control the level of residual noise and make a trade-off between variance and deviation. The experimental results show the proposed method achieved good denoising effect. Compared with some other excellent methods, the proposed method increased the peak signal-to-noise ratio (PSNR) by about 1.0-3.2 dB for seismic images without discontinuity, and about 1.8-3.9 dB for seismic images with discontinuity.