Color treatment in endoscopic image classification using multi-scale local color vector patterns

In this work we propose a novel multi-scale operator which is based on the full color information within an image. In order to evaluate the method, we extract features from endoscopic images using this operator and classify the images according to the respective class of polyps. [Display omitted] ► Compared to other LBP-based operators LCVP uses all color information available, yet yielding a more compact descriptor for an image. ► LCVP is up to 7.5 times faster compared to other LBP-based methods evaluated. ► In terms of a classification of polyps the accuracy of LCVP differs insignificantly only from previously developed methods. In this work we propose a novel method to describe local texture properties within color images with the aim of automated classification of endoscopic images. In contrast to comparable Local Binary Patterns operator approaches, where the respective texture operator is almost always applied to each color channel separately, we construct a color vector field from an image. Based on this field the proposed operator computes the similarity between neighboring pixels. The resulting image descriptor is a compact 1D-histogram which we use for a classification using the k-nearest neighbors classifier. To show the usability of this operator we use it to classify magnification-endoscopic images according to the pit pattern classification scheme. Apart from that, we also show that compared to previously proposed operators we are not only able to get competitive classification results in our application scenario, but that the proposed operator is also able to outperform the other methods either in terms of speed, feature compactness, or both.