The most popular image classification framework consists of two major modules: BoW and SPM. The framework of SPM based on BoW has been successfully applied to image classification [1,2] and in recent years it has been improved for classifying staining patterns of HEp-2 cells [3, 4]. It seems to be suitable for the staining pattern classification task.
There are four basic steps in the framework used for image classification as illustrated in Fig. 4.1. Each step within the framework can affect the quality of image representation and the following classification performance. The basic steps are respectively:
© Springer International Publishing AG 2017 X. Xu et al., Cellular Image Classification, DOI 10.1007/978-3-319-47629-2_4
Fig. 4.1 Overview of the feature extraction-coding-pooling scheme
- • The image patches are sampled from the input images in a dense (e.g. using fixed and overlapped grids) or sparse (e.g. using feature extractors) manner. Local features are extracted within image patches. There are various features have been applied in this procedure. For example, Scale Invariant Feature Transform (SIFT) feature  is one of the most popular features, which describes a patch with the local accumulation of the magnitude of pixel gradients in each orientation. Other widely employed features include histogram of oriented gradients , local binary pattern , etc.
- • A codebook is generated based on the features extracted from all training images. Usually, the codebook is generated by clustering methods (e.g. k-means) or learned in a supervised [8, 9] or an unsupervised [10, 11] manner.
- • Each feature actives a number of codewords in the codebook and is transformed as a coding vector, whose length is equal to the number of codewords. There are various coding algorithms to generate coding vector in different ways. How to encode the features is an essential procedure within the BoW framework.
- • To capture the shapes or locating an object, SPM  is proposed by dividing the image into increasingly finer spatial subregions. The final image representation is generated by integrating histogram of mid-level features (codes for local features) from each subregion. Typical pooling methods include max-pooling and averagepooling .
Feature coding and pooling has great impact on the following image classification in terms of both accuracy and efficiency. In this chapter, we briefly introduce the coding and pooling procedure.