GHOST ARTIFACT DETECTION AND REMOVAL IN HDR IMAGE PROCESSING USING MULTI-SCALE NORMALIZED CROSS-CORRELATION

摘要

A method for detecting ghost artifact in multiple images during high dynamic range (HDR image creation. For each pair of images, multiple individual consistency maps are generated by calculating a consistency function using moving windows of different sizes. The individual consistency maps are combined into a combined consistency map and then binarized. When combining the multiple individual consistency maps, they are first binarized using predetermined threshold values that are window-size-specific. The threshold values are developed beforehand using training images and machine leaning. The final binarized consistency map indicates whether the pair of images are consistent with each other at each pixel location. Then, a ghost-weight map is generated for each image based on the multiple final consistency maps. The ghost-weight map indicates the likelihood of each image pixel being ghost-inducing. The HDR image is generated using the set of images and the corresponding ghost-weight maps.

主权项

1. A method for generating a high dynamic range (HDR) image from a set of multiple images at different exposure levels, comprising:
for each pair of images of the set of images:
(a) generating a plurality of individual consistency maps for the pair of images, each individual consistency map being generated by calculating a consistency function between a pair of moving windows having a window size and located respectively in the pair of images, the plurality of individual consistency maps being generated using a plurality of different window sizes;(b) combining the plurality of individual consistency maps to generate a combined consistency map for the pair of images; and(c) binarizing the combined consistency map to generate a final consistency map for the pair of images, wherein each pixel in the final consistency map has a pixel value indicating whether or not the pair of images are consistent with each other at a location corresponding to that pixel; whereby a plurality of final consistency maps are generated, one for each pair of images;
(d) for each image of the set of images, generating a ghost-weight map based on the plurality of final consistency maps, wherein each pixel in the ghost-weight map has a pixel value indicating a likelihood of an image pixel at a corresponding location of the image being ghost-inducing; and(e) generating the HDR image using the set of images and the corresponding ghost-weight maps.