Research On Image Feature Extraction Algorithm Based On Geometric Algebra

In recent years,researches on classification and recognition of images have become more and more attractive in the field of computer vision and pattern recognition.As the basis of image classification and recognition,the feature extraction algorithm is favored by many researchers.In this paper,geometric algebra is used as a mathematical tool,combined with the characteristics of video images and hyperspectral images,in-depth study on the key techniques of video image local invariant features and depth feature extraction algorithms and in hyperspectral image local invariant feature extraction algorithms.The main innovative work as follow:(1)The motion information of the video is important especially when the target has a large motion.Besides,the motion information plays a key role in the representation of the video.And it is another key information except the apparent information.In order to make full use of the apparence information in the spatial domain and motion information in the temporal domain of the video,this paper uses geometric algebra as the mathematical framework to construct a spatio-temporal geometric algebra unified model of apparence and motion-variation(UMAMV)information for video.(2)Hyperspectral image spectral domain information is important for reflecting material properties.This paper uses the gradient information on the spectral domain and apparence information on the spatial domain to represent the macro information contained in the hyperspectral image.Base on the mathematical framework of geometric algebra,we construct a spatial-spectral geometric algebra unified model of spectral value and gradient change(UMSGC)for hyperspectral images.(3)A UAMMV-SIFT feature detector based on the UMAMV of video is proposed.The local invariant feature of video is the basis for pattern classification of video,while Scale-invariant feature transform(SIFT)has good resistance to light changes,occlusion,scale,size changes,the change of perspective affine transformation and the influence of noise.The UMAMV-SIFT feature detection algorithm first constructs the scale space of the UMAMV of the video,and then constructs the Gaussian pyramid and the difference of Gaussians(DoG)filter by means of variable scale.Finally,thefeature points are detected by the conducting non-maximum suppression on the DoG.The experimental results show that the proposed algorithm can detect the feature points with obvious motion more effectively than the traditional feature extraction method,and can extract more local invariant features with more accurate positions.(4)A UMAMV-TSN network based on video UMAMV is proposed.The temporal segment networks(TSN)modeling video is based on the long-term temporal structure.It is combined with the sparse sampling strategy in the temporal domain and the video-level supervision to make the whole video learning more effective and efficient.First,the UMAMV-TSN network segments and samples a video.Second,a UAMMV model is constructed containing the apparent information and motion information for each video segment.And then the above information is fed into the spatial network and the time network respectively.And use the segment consensus function to obtain a consensus on the category hypotheses between the segments,and then make a category prediction for the whole video.Finally,the video is classified and identified by fusing two networks with a certain weight ratio.The experimental results on UCF101 and HMDB51 datasets verified the effectiveness of the UMAMV-TSN network.(5)The UMSGC-SIFT feature extraction algorithm based on hyperspectral images UMSGC is proposed.During the feature detection process,the scale space of the UMSGC is generated firstly.And then the Gaussian pyramid and the difference of Gaussians are constructed on the scale space.The local invariant features are searched on the DoG,which reflects the spectral value and the gradient change information.In order to fully reflect the shape and texture characteristics of the local image near the local features and the material properties of the spectral domain,we propose a new UMSGC-SIFT feature descriptor to describe the extracted local features after extracting the local features.The experimental results show the effectiveness and accuracy of the UMSGC-SIFT algorithm in feature extraction of hyperspectral images.