Desktop version

Home arrow Engineering arrow Cellular Image Classification

Experiments and Analyses

In this section, we conduct a series of experiments to validate the efficiency of our proposed method for staining pattern classification on two HEp-2 cell datasets: the ICPR2012 dataset and the ICIP2013 training dataset as described in Sect. 1.3.

Experiment Setup

We quantitatively compare the classification performance achieved by our proposed method, AdaCoDT, with several LBP related features, namely, conventional LBP [1], CoALBP [19] (the winner of ICPR HEp-2 cell classification contest) and RICLBP [20]. As our proposed AdaCoDT method is a generative version of the BoW framework, we also compare it with BoW representation. LSC algorithm is chosen in our experiments due to its computationally efficiency and superior performance for staining pattern classification [9, 10, 24]. In order to make the evaluation comprehensive, we also use the FK based on dense SIFT (FK-SIFT) and the LSC based on CoDT (LSC-CoDT) as comparison. The protocols are used as the ICPR’12 HEp-2 cells classification contest and the parameters are optimized manually via several trials. The parameters for the comparative methods are set as Table7.1. Two parameters need to be considered while extracting LBP related features, that is, the number of neighbor pixels, P, and the radius, R. The interval between the LBP pair, d, should also be taken into account for CoALBP and RIC-LBP. The number of GMM components T is another parameter to be considered for the FK based methods. With respect to the LSC method, the codebook size is chosen as 1024 due to the trade-off between classification accuracy and computational cost. The parameters of the proposed AdaCoDT method will be discussed in Sect. 7.3.4.

Table 7.1 Parameters for comparative algorithms


(P, R) or (P, R, d)






(4, 1,2), (4, 2, 4), (4, 4, 8)



(4, 1,2), (4, 2, 4), (4, 4, 8)



(8, 1), (12, 2), (16, 3)


< Prev   CONTENTS   Source   Next >

Related topics