Improved library shelf reading using color feature matching of book-spine images

Many computer vision algorithms for pattern matching, object tracking, and 3-D reconstruction, etc., begin with feature detection and matching. Common feature detectors such as Harris, Sobel, Canny, and Difference of Gaussians perform basic linear algebra operations on an image in order to identify "corners" or "edges" for matching. These detectors however, require single-channel (grayscale) source images. For color source images, color channels are typically averaged or converted to create grayscale images for processing, discarding a large amount of highly useful information. This paper outlines the proposed color Difference of Gaussians (DoG) algorithm for feature detection, which outperforms the grayscale DoG in number and quality of features found. The new color DoG is applied to a color feature matching application for improving the library inventory (shelf-reading) process. Experimental results demonstrate the robustness of this color feature detector.