Research On Feature Extraction And Classification Method Of RGB-D Images

Image classification is always a challenging and difficult research problem in the field of computer vision and pattern recognition.It has been widely used in many important fields,such as security problems,remote sensing monitoring,image retrieval and so on.However,the classic image classification algorithms are mostly based on RGB or grayscale images,and the depth information of the object or scene has not been utilized effectively.Depth information has been a very important clue of image classification,can directly reflect the characteristics of 3D object or scene.With the rise of Kinect sensor,it becomes easier to accept depth image.And it has been a hot area of research to classify Image combining depth information in the field of computer vision and pattern recognition.This paper majors in research of how to improve the accuracy of image classification combining depth information.In view of the traditional image classification algorithm based on RGB-D,the feature of image extraction is too single,the resulting fusion features can't fully represent the image content.Based on the superiority of the integrated algorithm,this paper proposes an image classification approach which integrates the RGB-D fusion feature.The algorithm integrates four different fusion features by using the dense image of Dense Scale Invariant Feature Transform(dense SIFT)and histogram of Oriented Gradient(HOG)on the RGB image and depth image.K-means clustering algorithm to establish four different visual vocabulary,then combined with local-constrained linear coding(LLC)to characterize them to obtain a different set of image representation.Finally,on the basis of this,we use linear support vector machine(SVM)to classify and use voting decision method to determine the final classification.The traditional descriptor based on gradient feature does not apply to feature extraction of depth image.In view of this,this paper proposes a feature extraction algorithm based on the depth image,and applies it to the RGB-D image classification,this paper proposes an image classification method based on RGB-D fusion feature.Firstly,the dense SIFT feature of color image is fused with the global Gist feature of the depth image to generate a combined vector.Secondly,the improved k-means algorithm is used to build the visual dictionary of the fusion feature,which is to overcome the dependence on the initial point selection of traditional K-means algorithm.Moreover,in the stage of image representation,the approximate LLCfeature coding method is introduced to operate sparse coding on feature base and its corresponding visual dictionary.Finally,the linear SVM is used for image classification.In order to verify the effectiveness of the proposed algorithms,experiments are carried out on the open RGB-D Object and RGB-D Scenes data sets.The experimental results show that the proposed algorithms have good accuracy and robustness.