Does M-Net Can Achieve A Satisfactory Result On The SOC Experiment?

From TABLE V, we can see that the overall test accuracy under SOC reaches 99.71% and only 7 test samples are misclassed. The accuracies of per class are all more than 99%.
Especially BTR-70, T-72, T62, ZSU234 four classes even achieve 100%. The results in TABLE 5 show that M-Net can really achieve a satisfactory result on the SOC experiment.
F. Results and analysis of EOC experiment
The EOC experiment is divided into three partial experiments for using different data, namely EOC-1, EOC-2 and EOC-3.
In EOC-1, four kinds of data, 2S1, BRDM-2, T-72, and ZSU-234, are tested. Since MSTAR dataset is limited, only these four kinds of data can be used in this experiment. So, EOC-1 is a four-target classification problem. EOC-1 has bigger difference
This shortcoming makes these algorithms can’t exert their true power and it is hard to get a satisfactory result in the practical application. In order to analyze M-Net’s sensitivity to the variation of training samples number, M-Net performs accuracy tests on SOC experiment with the changeable number datasets. The test data used in this experiment is the same as the test data in TABLE IV and the training data is randomly selected from the training data in TABLE IV. For getting an authentic and precise result, this experiment will not use data argument method. The relationship between probability of misclassification and the number of training samples is described as shown in Fig.
6. Relationship between training sample number and EOC error rate.
It can be seen from Fig. 4 that M-Net's error rate curve is always below the other three algorithms’ curves and has a much more graceful decay. The other curves suffer a severe degradation when the training sample number changes. What is indicated is that M-Net gets the highest accuracy in every training sample number in the graph. When the training sample number is greater than 100, M-Net’s curve becomes flat and smooth. It shows that when the training sample number reaches a sufficient level, the changing of it has insignificant influence on the M-Net. Thus, M-Net can get an outstanding and stable performance in a wide range.
H. Feature extraction
In the SOC experiment, the total number of pictures in the test set is 2425. Each picture can be extracted a feature vector with Nu dimensions by M-Net. In order to visualize the feature vectors, Principal Component Analysis (PCA) technique is adopted to reduce the dimensionality of feature vector from Nu dimensions to two dimensions [43] which can be depicted in a two-dimensional graph, as shown in Fig. 7. (a)