Joint Sparse Coding Spatial Pyramid Matching for Classification of Color Blood Cell Image

Abstract

Abstract. In the Automatic recognition of blood cell images, the color blood cell images are usually transformed into grayscale images for feature extraction, which result in losing plenty of useful color information. Although the sparse coding based linear spatial pyramid matching (ScSPM) is popular in grayscale image classification, the sparse coding methods in ScSPM fail to extract color information. In this paper, we proposed a novel joint sparse coding SPM (JScSPM) method by using the joint trained joint codebook. The joint codebook is able to represent the inner color correlation among different color components, and the individual color information of each color channel as well. JScSPM method was then applied to classify color blood cell images. The experimental results showed that the proposed method achieved mean 3.1% and 6.6% improvements on classification accuracy, compared with the majority voting based ScSPM the original ScSPM, respectively. IntroductionThe analysis of blood cells in microscope images is one of the most important steps in clinical hematological procedures. Consequently, the image-based automatic counting and classification of blood cells have attracted lots of attention. In current classification system, the colored blood cell images are usually transformed into grayscale images for further feature extraction, in order to simplify algorithms and reduce the computational complexity. However, plenty of useful color information will be discarded. Furthermore, there are few special designated color feature descriptors for blood cell images.In recent years, the sparse coding technique has been successfully used in image classification task [1], not only directly as a classifier, but also embedded in the classification framework [2]. The sparse coding based linear spatial pyramid matching (ScSPM) is a popular sparse coding embedded classification method developed on spatial pyramid matching (SPM) [2,3]. It computes a spatial pyramid image representation with sparse coding instead the vector quantization [2,3]. The ScSPM and its variations [2,4,5] have achieved great success in image classification. The main improvements in ScSPM mainly focus on applying different regularization terms for