*5.3. E*ff*ect and Stability Analysis*

The train number of each species (*tr\_no*) was changed from 5 to 30 as shown in Figure 12. To each training set, SIFT with LLC coding was used for comparison, and the training set and testing set were kept the same for each feature. All the accuracies are the average of 10 times.

**Figure 12.** The relationship between the training number of each species and the recognition accuracy.

It is obvious that, for each feature, the accuracy increases with *tr\_no*. However, we are most concerned with the proposed feature BOF\_DP showing a better effect than BOF\_SIFT. Both BOF\_SIFT and BOF\_DP are better than BOF\_SC, and the combined features of BOF\_DP and BOF\_SC achieved the highest recognition accuracy in the Flavia dataset.

To show the results of recognition clearly, recognition rates for each species are shown in Figure 13. The training number of each class was 30, and the final recognition rate on Flavia dataset was 98.2049%. Except for Species 11 and 25, the recognition accuracies of other species were ideal.

#### *5.4. Comparison of Features*

Some other features were used for comparison, as shown in Table 1. BOF\_DP represents the proposed feature, BOW + SIFT represents the features in Ref. [14], BOW + SC is also a proposed method based on SC and BOW in Ref. [14], LLC + SIFT is the original LLC method using SIFT [29], DBCS is a deformation-based representation space for curved shapes, and the authors of [39] proposed an adaptation of k-means clustering for shape analysis in DBCS. 2DPCA [40] is the 2D-based method of principal component analysis (PCA) and uses the bagging classifier with the decision tree as a weak learner. The recognition accuracies of these features are relatively close. 2DPCA has the lowest accuracy among these features. The proposed feature BOF\_DP obtains the highest accuracy in the comparison.


**Table 1.** Comparison of proposed feature with existing features on Flavia dataset.
