In order to capture the low-level visual features of different scenes accurately, we
can find some discriminating key points in the image. Such as corner point, edge,
center point of specific area, endpoint, center of gravity, extreme point, intersection,
etc.) [6]. The most common key point is the corner point, which corresponds to the
dramatic changes in image brightness or curvature maximum point in edge curve of
image. The corner of a large number of features, such as road intersection in remote
sensing images is presented as corner points. The neighborhood of the key points is
also a region which has stable and rich information in the image. Because the difference
of geometric relationship between remote sensing image and natural image is large,
different imaging viewpoint often leads same scene in remote sensing images to appear
larger difference. So it is necessary to search key point feature extraction algorithm
which maintains a certain invariance to image changes such as rotation, scaling, affine
transformation, angle of view, illumination changes and so on and maintains good
matching with the object motion, cover, noise, and other factors.
The feature extraction algorithm SIFT (Scale Invariant Feature Transform) proposed by David G. Lowe not only has invariance to the scale and rotation of the image
but also has certain stability to the brightness and the angle of view. So key points of
the image can be detected effectively by applying SIFT algorithm. Because the scale
range of remote sensing image has larger variation, and the scene types are complex
and diverse, so it is obviously low efficiency if segment region first and then extract
feature. And local invariant feature points are usually sparse. Although they have good
effect on the natural images, it is not robust enough for the complex and changeable
remote sensing image.
The classical SIFT operator is effective for grayscale images, Considering the
spectral feature is an important sign to distinguish the type of remote sensing image
scene, this paper introduced the spectral information, which can help to improve the
performance of scene recognition based on visual feature representation. Firstly, the
normalized multi-spectral remote sensing image is transformed into HSV space, and
the image is divided into a dense grid of n * n, SIFT feature description of the grid
center points are calculated on the HSV channels. And Splice them together into a 384dimensional vector as CSIFT characteristics of the grid. Do K-means clustering of
CSIFT features extracted from all the training images. And M visual words corresponding to M clustering center is obtained.
For an under-recognized remote sensing image (or region), the CSIFT feature is
extracted by the same method as the training image. According to the nearest neighbor
rule, determine the visual words classification of each CSIFT. Count the frequency of
the visual words in the remote sensing image (or region) to be classified, and obtain
visual bag of features of the under-recognized remote sensing image. The visual bag of
feature of remote sensing image avoids by the process of target segmentation and
detection in the scene, and transforms the identification of the scene into the learning
problem of visual word distribution. So build a bridge for “the semantic gap” between
low-level features and high-level semantics.
380
Y. Sun et al.
can find some discriminating key points in the image. Such as corner point, edge,
center point of specific area, endpoint, center of gravity, extreme point, intersection,
etc.) [6]. The most common key point is the corner point, which corresponds to the
dramatic changes in image brightness or curvature maximum point in edge curve of
image. The corner of a large number of features, such as road intersection in remote
sensing images is presented as corner points. The neighborhood of the key points is
also a region which has stable and rich information in the image. Because the difference
of geometric relationship between remote sensing image and natural image is large,
different imaging viewpoint often leads same scene in remote sensing images to appear
larger difference. So it is necessary to search key point feature extraction algorithm
which maintains a certain invariance to image changes such as rotation, scaling, affine
transformation, angle of view, illumination changes and so on and maintains good
matching with the object motion, cover, noise, and other factors.
The feature extraction algorithm SIFT (Scale Invariant Feature Transform) proposed by David G. Lowe not only has invariance to the scale and rotation of the image
but also has certain stability to the brightness and the angle of view. So key points of
the image can be detected effectively by applying SIFT algorithm. Because the scale
range of remote sensing image has larger variation, and the scene types are complex
and diverse, so it is obviously low efficiency if segment region first and then extract
feature. And local invariant feature points are usually sparse. Although they have good
effect on the natural images, it is not robust enough for the complex and changeable
remote sensing image.
The classical SIFT operator is effective for grayscale images, Considering the
spectral feature is an important sign to distinguish the type of remote sensing image
scene, this paper introduced the spectral information, which can help to improve the
performance of scene recognition based on visual feature representation. Firstly, the
normalized multi-spectral remote sensing image is transformed into HSV space, and
the image is divided into a dense grid of n * n, SIFT feature description of the grid
center points are calculated on the HSV channels. And Splice them together into a 384dimensional vector as CSIFT characteristics of the grid. Do K-means clustering of
CSIFT features extracted from all the training images. And M visual words corresponding to M clustering center is obtained.
For an under-recognized remote sensing image (or region), the CSIFT feature is
extracted by the same method as the training image. According to the nearest neighbor
rule, determine the visual words classification of each CSIFT. Count the frequency of
the visual words in the remote sensing image (or region) to be classified, and obtain
visual bag of features of the under-recognized remote sensing image. The visual bag of
feature of remote sensing image avoids by the process of target segmentation and
detection in the scene, and transforms the identification of the scene into the learning
problem of visual word distribution. So build a bridge for “the semantic gap” between
low-level features and high-level semantics.
380
Y. Sun et al.