Defining Feature Space for Image Classification
Abstract In this chapter, we design a Co-occurrence Differential Texton (CoDT) feature to represent the local image patches of HEp-2 cell image, and a generative model is built to adaptively characterize the CoDT feature space. We further exploit a more discriminant representation for the HEp-2 cell images based on the adaptive partitioned feature space, and then feed the representation into a linear SVM classifier to identify the staining patterns. Two benchmark datasets are used for evaluation on the classification performance of our proposed method.
It is verified that powerful discrimination can be achieved by exploiting the distributions of gray values over extremely compact neighborhoods (starting from as small as 3 x 3 pixels) [1,2]. LBP related features have been applied successfully in the staining pattern classification [3, 4]. The main idea is to characterize each microstructure into a binary series by thresholding the gray value of the neighboring pixels by that of the center. However, some important information is lost since the LBP represents the local structures with only two quantized levels, i.e., 0 and 1 . In this study, we propose a CoDT feature without quantizing over micro-structure to characterize the HEp-2 cells. Furthermore, we capture the spatial relations among the differential micro-texton features to improve the discriminative power of features.
Next, the BoW framework, which is very influential in image classification, has been proposed as a set of visual words selected from a codebook learned beforehand [6, 7]. As we mentioned in Chap.5, it has been applied in classification of staining patterns and has shown impressive performance [8-11]. The BoW framework seems suitable for the purpose of staining pattern classification. However, the BoW approach suffers from its own problems to which we would like to explore for the solution: (i) Information loss in feature quantization process is inevitable and affects the performance for good image classification ; (ii) The computational cost of histogram generation depends directly on the size of codebook. Since better performance is always obtained with larger vocabulary, the computational cost of
© Springer International Publishing AG 2017 X. Xu et al., Cellular Image Classification, DOI 10.1007/978-3-319-47629-2_7
the BoW framework is high. (iii) The reason for such a histogram representation has to be optimal for the specific classification task is in question .
To handle these problems, we make use of the Fisher Kernel (FK) principal [14, 15] for staining pattern classification to avoid aforementioned problems of the BoW framework. It characterizes an image by a gradient vector derived from a generative model of the training data, thus the resulting representation is adapted to the classification task. Meanwhile, with the same size of vocabulary, the FK based method can extract much larger image representation than the BoW representation. Hence, it can provide excellent performance with simple linear classifier. For staining pattern classification of HEp-2 cells, we apply a Gaussian Mixture Model (GMM) to adaptively approximate the distribution of our proposed CoDT features. Then we can obtain a high discriminative and powerful descriptive HEp-2 cell image representation, which is intrinsically adapted to the classification task, based on the adaptive CoDT feature space.
Our major contributions are threefold: (i) A local feature, CoDT, is defined and extracted directly using the differential vectors of micro-texton and its neighborhoods. The information loss caused by quantization of the LBP related features is reduced, and the spatial information is captured simultaneously. Therefore our proposed CoDT feature provides powerful discriminative capability; (ii) The CoDT feature space is adaptively characterized as a generative model, GMM. Thus the parameters can be adjusted from the training cell images which are better fitting the CoDT feature space; (iii) The final image representation, which is derived from the generative model, can cooperate with a simple linear SVM classifier for identifying staining pattern of HEp-2 cells. Our proposed framework (AdaCoDT) can exploit the advantages of both generative and discriminative approaches for image classification. Experimental results verify that our proposed method can provide much better performance of staining pattern classification than that of the traditional LBP and its extensions. It also improves the classification performance in comparison of the BoW representation.