Desktop version

Home arrow Engineering arrow Cellular Image Classification

Discussion

To further analyze our proposed AdaCoDT method, we evaluate its performance with respect to the CoDT size (P, R, d) and the number of GMM components T. It should be noted that, the classification performance mentioned in this section means average accuracy at the cell level.

The CoDT size (P, R, d): we study the effect of different parameter (P, R, d) for classification performance. As shown in the Table7.8, the best classification performance is achieved using parameters (24,4, 8) and (16, 5,10) for ICPR2012 dataset and ICIP2013 training dataset respectively.

Table 7.8 Classification performance of the AdaCoDT method with different parameters (P, R, d)

(8,1,2)

(16,3,6)

(16,4,8)

(16, 5, 10)

(24, 3, 6)

Acc (ICPR2012) (%)

53.9

70.1

72.9

71.7

72.9

Acc (ICIP2013) (%)

64.0

73.1

74.2

75.8

73.1

(24, 4, 8)

(24, 5, 10)

(32, 4, 8)

(32, 5, 10)

(32, 6, 12)

Acc (ICPR2012) (%)

75.2

73.2

74.8

74.1

70.8

Acc (ICIP2013) (%)

74.4

74.4

73.6

73.2

73.0

Classification performance of the AdaCoDT method under different number of GMM components

Fig. 7.4 Classification performance of the AdaCoDT method under different number of GMM components

The number of GMM components: we test the performance under various number of GMM components. Figure7.4 presents the cell-level classification performance of the AdaCoDT method obtained with respect to increasing number of GMM components. As can be seen, the AdaCoDT method with larger number of components can generally achieve better performance for both HEp-2 cells datasets. The classification performance is steady from 128 to 256 for ICPR2012 dataset while the classification accuracy of ICIP2013 training dataset is slightly declining. In our experiments, we choose T = 256 for ICPR2012 dataset and T = 128 for ICIP2013 training dataset due to the trade-off between accuracy and memory usage.

 
Source
< Prev   CONTENTS   Source   Next >

Related topics