Patch Normalization for Visual Template Matching based on Panoramic Images in Bio-inspired Navigation System

The excerpt note is about how to enhance the signal-to-noise ratio by using patch normalization in each image to enhance edge information and eliminate image intensity variation from Edward P. et al. IJRR 2016.

Patch normalization is performed by dividing an image into a grid of square patches and for each pixel, subtracting the patch mean and then dividing by the patch standard deviation. Finally, a mean difference score, , is calculated for each query-database image pair by calculating the -norm with the Sum of Absolute Difference (SAD) metric

Where returns a central region of the image, and returns a region of the image, offset by from the center, where and , and and are the dimensions of the low-resolution, patch-normalized grayscale images. We compare central subregions of and over a range of offsets up to horizontal and vertical maxima ( and , respectively), such that the SAD score of the overlapping region is minimized. As each query frame is compare with all database frames using SAD, we form a difference vector, , for each frame.

Lastly, we modify the raw difference scores by considering that we are searching for coherent sequences of locally-best image matches.

Hence, the local matching contrast is enhanced by normalizing each element in the difference vector within a neighbourhood centered around it

Where and are the mean and standard deviation of the neighbourhood vector, respectively

Where is defined by a neighbourhood radius parameter, , and bounded by the length of . This process gives the normalized difference vector, .

References:

Pepperell, Edward, Peter Corke, and Michael Milford. “Routed roads: Probabilistic vision-based place recognition for changing conditions, split streets and varied viewpoints.” The International Journal of Robotics Research 35, no. 9 (2016): 1057-1179.