Patch-Based Image Processing: from Dictionary Learning to Structural Clustering

1.1 Historical Background and Overview Where does patch come from? According to Publish or Perish, the earliest appearance of " patch " as a technical term in the open literature was in Haralick's 1973 paper on textural features for image classification [1] (the term could exist before that but it is difficult to track back further). In those old days, pixels are termed " " and small-area typically refer to a window or patch of 3 × 3 pixels. Seven years later, in JS Lee's pioneering paper on image enhancement [2], local statistics (e.g., mean and variance) are calculated on a patch-by-patch basis in order to drive the filter to remove additive or multiplicative noise. In early 1980s, computing optical flow (e.g., Horn's method [3]) or disparity field (e.g., Lukas-Kanade algorithm [4]) from a pair of images is also based on the assumption that the gradient of image intensity field can be locally calculated within a patch where patch is often interchanged with the term window or block. Fast advances of communication and computing technologies in later 1980s and early 1990s stimulated the research into image compression especially the development of standard [5]. Under the context of JPEG compression,