Conclusions and Perspectives
Abstract In this last chapter, we concludes this monograph with its major techniques developed, and give our perspectives on the future directions of research in this field.
Major Techniques Developed in the Book
In this book, we aimed at improving the performance for classifying staining patterns of HEp-2 cells to support the diagnosis of specific autoimmune diseases. Especially, we focused on extracting suitable and effective features for representing the HEp- 2 cells w.r.t the staining patterns. We explored three kinds of image descriptors, from low-level local textural descriptor to mid-level coding feature. Experimental evaluations on two publicly available HEp-2 cells datasets validate the effectiveness of our methods.
Firstly, we researched on the BoW framework which had been applied successfully in the staining pattern classification. We improved the coding method and proposed a promising framework, LLDC, for automatic staining pattern classification. Our proposed method is based on local distance vector which can capture discriminative information. Furthermore, it improves the classification performance by making adjustments only to the classes found in the local few nearest neighbors around the local features. To further improve the classification accuracy, two image representations are concatenated together as the distance patterns and the original local features are proven to be complementary to each other.
Then, we integrated the characteristic of rotation invariant into the textural features. We proposed to extract two kinds of rotation invariant descriptors: the PLTP- SRI feature and the BoW representation based on dense SIFT. Our proposed method takes the advantages of the two kinds of features from different aspects. It has the advantages of invariance under image rotations, meanwhile it has strong discriminative and descriptive ability. Incorporated with a linear SVM classifier, our method demonstrates its effectiveness by experimental results consistently.
Thirdly, we designed CoDT features to represent HEp-2 cell images. 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. Then, we built a generative model
© Springer International Publishing AG 2017 X. Xu et al., Cellular Image Classification, DOI 10.1007/978-3-319-47629-2_8
to adaptively characterize the CoDT feature space. We further exploited a more discriminant representation for the HEp-2 cell images based on the adaptive partitioned feature space, and fed the representation into a linear SVM classifier for identifying the staining patterns. Our proposed framework (AdaCoDT) can exploit the advantages of both generative and discriminative approaches for image classification.