Identifying Recurring Patterns with Deep Neural Networks for Natural Image Denoising
share
While there is a vast diversity in the patterns and textures that occur across different varieties of natural images, the variance of such patterns within a single image is far more limited. A variety of traditional methods have exploited this self-similarity or recurrence with considerable success for image modeling, estimation, and restoration. A key challenge, however, is in accurately identifying recurring patterns within degraded image observations. This work proposes a new method for natural image denoising, that trains a deep neural network to determine whether noisy patches share common underlying patterns. Specifically, given a pair of noisy patches, the network predicts whether different transform sub-band coefficients of the original noise-free patches are the same. The denoising algorithm averages these matched coefficients to obtain an initial estimate of the clean image, with much higher quality than traditional approaches. This estimate is then refined with a second post-processing network, yielding state-of-the-art denoising performance.
READ FULL TEXT
