4 Semantic Recognition of Scenes Based on PLSA
Although the frequency of visual words can be used as an important basis for distinguishing different scenes, in a complex remote sensing image scene, the same target
entity may appear in different scene categories. So it may lead to ambiguity and
similarity problems between visual words and scene semantics. In the case of insufficient training samples, by using an identification method that associates the scene
categories with the extracted feature vector directly, it cannot approximate the actual
scene semantics and will reduce accuracy of scene recognition. Our thinking is to apply
the probabilistic latent semantic analysis model to typical training image, to extract the
latent semantic of image. And according to the probability distribution of latent
semantic, determine the scene type of identified images.
(1) Extract the characteristics of all images. Selected part of image randomly from
each class of training image, extract CSIFT feature vector of these images, and
generate M visual vocabularies by k-means clustering algorithm. To measure the
similarity of each visual word with the feature vector of each training image, and
obtain “image–Words” co-occurrence frequency matrix n 0 (d i , w j ) which is
N * M dimension. Among them i 2 (1, N), j 2 (1, M). It represents frequency of
visual words w j in image d i .
(2) Obtain the approximate maximum likelihood estimation result of PLSA model by
using EM algorithm, And then obtain distribution pðw j jz Train Þ of visual words
when latent semantic appear in the image.
(3) Extract the feature vector of the test image. A similar measure respectively with M
visual vocabularies obtained in (1). Obtain “image–Words” co-occurrence frequency matrix n T (d T , w j ) of test image. The pðw j jz Train Þ and co-occurrence frequency matrix n T (d T , w j ) of test image is used as the input of the PLSA model,
remain pðw j jz Train Þ unchanged. The latent semantic distribution p(z | d T ) of the test
image is obtained. And K dimensional semantic vector of the test image is built.
(4) The KNN classifier is applied to the latent semantic vector of the test image, and
the scene recognition of image is accomplished.
5 Experimental Results and Analysis
Total 1794 image slices from 10 different type scenes are acquired from the Earth
Google. The resolution and size of the image are not limited. The scene type is
determined by the main object of the slice. 50 images are selected randomly from the
10 types of images as training images and the rest as testing images.
In order to verify the performance of the algorithm in this paper, firstly, we analyze
the effect of different feature extraction methods with the recognition results, further
compare the recognition effect by nearest neighbor classification between BOF
expression based on CSIFT features directly and introducing PLSA model. Finally,
compare the recognition results with different visual words number and different latent
semantic topics.
A Scene Semantic Recognition Method of Remote Sensing Images …
381
Although the frequency of visual words can be used as an important basis for distinguishing different scenes, in a complex remote sensing image scene, the same target
entity may appear in different scene categories. So it may lead to ambiguity and
similarity problems between visual words and scene semantics. In the case of insufficient training samples, by using an identification method that associates the scene
categories with the extracted feature vector directly, it cannot approximate the actual
scene semantics and will reduce accuracy of scene recognition. Our thinking is to apply
the probabilistic latent semantic analysis model to typical training image, to extract the
latent semantic of image. And according to the probability distribution of latent
semantic, determine the scene type of identified images.
(1) Extract the characteristics of all images. Selected part of image randomly from
each class of training image, extract CSIFT feature vector of these images, and
generate M visual vocabularies by k-means clustering algorithm. To measure the
similarity of each visual word with the feature vector of each training image, and
obtain “image–Words” co-occurrence frequency matrix n 0 (d i , w j ) which is
N * M dimension. Among them i 2 (1, N), j 2 (1, M). It represents frequency of
visual words w j in image d i .
(2) Obtain the approximate maximum likelihood estimation result of PLSA model by
using EM algorithm, And then obtain distribution pðw j jz Train Þ of visual words
when latent semantic appear in the image.
(3) Extract the feature vector of the test image. A similar measure respectively with M
visual vocabularies obtained in (1). Obtain “image–Words” co-occurrence frequency matrix n T (d T , w j ) of test image. The pðw j jz Train Þ and co-occurrence frequency matrix n T (d T , w j ) of test image is used as the input of the PLSA model,
remain pðw j jz Train Þ unchanged. The latent semantic distribution p(z | d T ) of the test
image is obtained. And K dimensional semantic vector of the test image is built.
(4) The KNN classifier is applied to the latent semantic vector of the test image, and
the scene recognition of image is accomplished.
5 Experimental Results and Analysis
Total 1794 image slices from 10 different type scenes are acquired from the Earth
Google. The resolution and size of the image are not limited. The scene type is
determined by the main object of the slice. 50 images are selected randomly from the
10 types of images as training images and the rest as testing images.
In order to verify the performance of the algorithm in this paper, firstly, we analyze
the effect of different feature extraction methods with the recognition results, further
compare the recognition effect by nearest neighbor classification between BOF
expression based on CSIFT features directly and introducing PLSA model. Finally,
compare the recognition results with different visual words number and different latent
semantic topics.
A Scene Semantic Recognition Method of Remote Sensing Images …
381