2 Schemes of Image Scene Recognition
At present, most of the scene recognition methods mainly focus on the basic level of
visual cognition. Firstly to extract the image features from the image, then use the
obtained image feature to design the semantic classifier, to implement mapping from
low-level features of image to high-level semantic. The main research methods include
two categories: low-level feature modeling method and middle semantic modeling
method.
There are mainly three types of methods in middle-level semantic modeling: [3],
(1) Construct the semantic object, to describe the scene by detecting or identifying
semantic objects in images. (2) Torralba and Oliva proposed Gist model, the model
avoids the segmentation of single object or region, to describe the scene structure by a
low dimensional space envelope. Among them, five kinds of sensory attributes, such as
nature, degree of openness, roughness, expansion, and roughness, separately corresponding to one dimension in space envelope. And each dimension as the basis of
semantic division of scene corresponds to a meaningful spatial attribute in the scene.
(3) Establish the local semantic concepts of image [4]. First, detect interest points
automatically in image, and describe these points by the local descriptor, then establish
the mapping of local descriptors to some local semantic concepts, realize the image
scene recognition by distribution of the local image semantic concept. This paper
mainly adopts this method for scene recognition of remote sensing images.
3 Visual Bag of Features of Remote Sensing Images
In order to realize the accurate recognition of remote sensing image scenes, no matter
adopt low-level feature modeling method or middle semantic modeling method, must
both extract discriminating features such as regional characteristics, sub-block features,
local invariant features from remote sensing images. Because different characteristics
reflect different types of information, it has its own advantages for a specific category.
And in many cases, the analysis of the image content also needs to combine different
characteristics. So the integration of multiple features is useful for improving the
performance of image scene recognition. Bag of Words [5] is the text simplified
description model used most commonly in text processing field. The model does not
consider the grammar and word order, only expresses the text into a disordered word
group. In text classified applications, often combined BOW model and the SVM
classifier, simple Bayesian classifier, obtained very good classification results [5]. After
the model is applied to the field of computer vision, it is popularized to Bag of Features
(BOF). Which basic principle is to generate visual words to describe the image or
image set by vector quantifying of various kinds local visual features.
The realization of BOF expression in remote sensing image is that detect the key
point of image firstly, then calculate the local descriptor around key points, quantize the
descriptor into visual words, and format visual words. Measure the frequency of the
specific visual words in the image, and finally establish the visual bag of features of the
image.
A Scene Semantic Recognition Method of Remote Sensing Images …
379
At present, most of the scene recognition methods mainly focus on the basic level of
visual cognition. Firstly to extract the image features from the image, then use the
obtained image feature to design the semantic classifier, to implement mapping from
low-level features of image to high-level semantic. The main research methods include
two categories: low-level feature modeling method and middle semantic modeling
method.
There are mainly three types of methods in middle-level semantic modeling: [3],
(1) Construct the semantic object, to describe the scene by detecting or identifying
semantic objects in images. (2) Torralba and Oliva proposed Gist model, the model
avoids the segmentation of single object or region, to describe the scene structure by a
low dimensional space envelope. Among them, five kinds of sensory attributes, such as
nature, degree of openness, roughness, expansion, and roughness, separately corresponding to one dimension in space envelope. And each dimension as the basis of
semantic division of scene corresponds to a meaningful spatial attribute in the scene.
(3) Establish the local semantic concepts of image [4]. First, detect interest points
automatically in image, and describe these points by the local descriptor, then establish
the mapping of local descriptors to some local semantic concepts, realize the image
scene recognition by distribution of the local image semantic concept. This paper
mainly adopts this method for scene recognition of remote sensing images.
3 Visual Bag of Features of Remote Sensing Images
In order to realize the accurate recognition of remote sensing image scenes, no matter
adopt low-level feature modeling method or middle semantic modeling method, must
both extract discriminating features such as regional characteristics, sub-block features,
local invariant features from remote sensing images. Because different characteristics
reflect different types of information, it has its own advantages for a specific category.
And in many cases, the analysis of the image content also needs to combine different
characteristics. So the integration of multiple features is useful for improving the
performance of image scene recognition. Bag of Words [5] is the text simplified
description model used most commonly in text processing field. The model does not
consider the grammar and word order, only expresses the text into a disordered word
group. In text classified applications, often combined BOW model and the SVM
classifier, simple Bayesian classifier, obtained very good classification results [5]. After
the model is applied to the field of computer vision, it is popularized to Bag of Features
(BOF). Which basic principle is to generate visual words to describe the image or
image set by vector quantifying of various kinds local visual features.
The realization of BOF expression in remote sensing image is that detect the key
point of image firstly, then calculate the local descriptor around key points, quantize the
descriptor into visual words, and format visual words. Measure the frequency of the
specific visual words in the image, and finally establish the visual bag of features of the
image.
A Scene Semantic Recognition Method of Remote Sensing Images …
379