5.3 Effect of Different Visual Words on the Recognition Results
In the previous experiment, the number of visual words is set to 600. Next just adjust
the number of visual words, sampling interval of dense grid and latent semantic number
is unchanged. Compare the final average recognition rate. The results are shown in
Fig. 2. The overall performance is relatively good when the number of visual words is
between 500 and 900. Less number of visual words will lead to a lack of discriminating
latent semantic topics. A large number of redundant information in the visual words
increases the ambiguity of the latent semantic topics, so it will lead to a decline in
overall performance.
5.4 Effect of Recognition Results Caused by Different Number of Latent
Semantic Topics
In the experiment, the number of the scene is 10, and the method of visual word
extraction is unchanged. The number of fixed visual words is 600, and the number of
latent semantic topics varies from 8 to 50. Compare the average recognition rate and
the result is shown in Fig. 3. The average recognition rate is the highest when the latent
semantic topic number is 20. Too many or too few potential semantic topics will lead to
decrease in the rate of recognition. Although the optimal number of visual words and
latent semantic topics exist in theory. It is very difficult to solve exactly in practical
application. It is mainly through a lot of experiments to determine an empirical value.
In order to further verify the feasibility of the method of scene semantic recognition, the recognition model is trained in the aforementioned experiments, and a remote
sensing image in North Korea Pyongyang area is used to make an interactive annotation. Taking into account the recognition model carry out specific targeted training
only for 10 kinds of typical scenes, so in the interactive selection, we try to select the
aforesaid 10 categories of scenes which can be manually identified. Identification result
shows that there are better recognition results in different types of scenes.
0
200
400
600
800
1000
1200
1400
1600
50
55
60
65
70
75
80
85
90
95
100
Number of Visual Features
Mean Recgnition Rate(%)
Fig. 2. The influence of different visual words numbers on recognition result
A Scene Semantic Recognition Method of Remote Sensing Images …
383
In the previous experiment, the number of visual words is set to 600. Next just adjust
the number of visual words, sampling interval of dense grid and latent semantic number
is unchanged. Compare the final average recognition rate. The results are shown in
Fig. 2. The overall performance is relatively good when the number of visual words is
between 500 and 900. Less number of visual words will lead to a lack of discriminating
latent semantic topics. A large number of redundant information in the visual words
increases the ambiguity of the latent semantic topics, so it will lead to a decline in
overall performance.
5.4 Effect of Recognition Results Caused by Different Number of Latent
Semantic Topics
In the experiment, the number of the scene is 10, and the method of visual word
extraction is unchanged. The number of fixed visual words is 600, and the number of
latent semantic topics varies from 8 to 50. Compare the average recognition rate and
the result is shown in Fig. 3. The average recognition rate is the highest when the latent
semantic topic number is 20. Too many or too few potential semantic topics will lead to
decrease in the rate of recognition. Although the optimal number of visual words and
latent semantic topics exist in theory. It is very difficult to solve exactly in practical
application. It is mainly through a lot of experiments to determine an empirical value.
In order to further verify the feasibility of the method of scene semantic recognition, the recognition model is trained in the aforementioned experiments, and a remote
sensing image in North Korea Pyongyang area is used to make an interactive annotation. Taking into account the recognition model carry out specific targeted training
only for 10 kinds of typical scenes, so in the interactive selection, we try to select the
aforesaid 10 categories of scenes which can be manually identified. Identification result
shows that there are better recognition results in different types of scenes.
0
200
400
600
800
1000
1200
1400
1600
50
55
60
65
70
75
80
85
90
95
100
Number of Visual Features
Mean Recgnition Rate(%)
Fig. 2. The influence of different visual words numbers on recognition result
A Scene Semantic Recognition Method of Remote Sensing Images …
383