Reseach On Image Super Resolution And Object Tracking Based On Sparse Representation

In recent years, computer vision has been developed rapidly, and it has a very wide range of applications in agriculture, industry, military, and so on. Computer vision in many ways can replace human vision to work. Image processing, object detection, recognition and tracking are the keys of the computer vision. The quality of the image directly affects on the further application. Super resolution method can help cameras to improve the image resolution with low cost, and provide more image details. Target detection, recognition, and tracking refer to identification in the images or videos for a target, so as to provide information for further decision making.Sparse representation(SR) has superior performance in image representation, and plays an important role in image super-resolution, target detection, recognition and tracking. SR has attracted wide attention from domestic and foreign researchers. In this thesis, the research work is carried out based on image super-resolution and target tracking via sparse representation. The main contents of the thesis include the following aspects:1. For the problem of low image resolution, a new method of image super resolution via group sparse representation(GS) is proposed, considering the image construction feature. First, an algorithm of combining the group orthogonal matching pursuit and K-SVD is proposed to train the dictionaries. The dictionary training can ensure that the corresponding low resolution(LR) and high resolution(HR) image have the same GS coefficients. Then the group sparse coefficients of the LR image are sought to reconstruct the HR image with the trained LR and HR dictionary pair. Experimental results indicate that the proposed method generates better results than several popular methods.2. A target detection method is proposed using sparse representation with element and construction combination(ECC) feature. First, the dense scale-invariant feature transform(SIFT) descriptors of image is extracted as the element features and correlations between each patch in the image are computed as the construction features. The element feather is local, and the construction feather is global. The two kinds of features are combined to represent the image. Second, the ECC feature is coded as sparse vector through a trained dictionary, and a feature histogram of sparse vector is computed based on spatial pyramid(SP). Then the feature histogram is fed into support vector machine(SVM) classifier. Finally the targets are detected in the activation map which is generated from the classifier. Experimental results demonstrate that the proposed method can detect targets with high performance, and it is robust to the target occlusion, for both the single scale target and the multi-scale target.3. A decision fusion method of SR and SVM is proposed based on Bayesian rule. First, a fast SR classifier(FSR-C) is proposed. In the FSR-C, the dictionary is composed of training images. Just one nonzero element in SR coefficient of the testing image is found based on matching pursuit, and the testing image is classified through the location of the nonzero element. Second, the SVM classifier(SVM-C) is introduced. In SVM-C, principal component analysis(PCA) feature is extracted, and to seek the linear separating hyperplane, the RBF kernel function is used in mapping the training vectors into high dimensional space. Finally, the results of the FSR-C and the SVM-C are fused obeying Bayesian rule to make the decision. The experimental results show that the proposed method improves the probability of classification accuracy and the stabilization.4. Object tracking methods based on PCA are effective against object change caused by illumination variation and motion blur. A new robust object tracking method based on the PCA and local sparse representation(LSR) is proposed. Firstly, candidates are reconstructed through the PCA subspace model in global manner. To handle occlusion, a patch-based similarity estimation strategy is proposed for the PCA subspace model. In the patch-based strategy, the PCA representation error map is divided into patches to estimate the similarity considering the occlusion. Secondly, the LSR is introduced to detect the occluded patches of object and estimate the similarity of candidates through the residual error in the sparse coding. Finally, the two similarities of each candidate from the PCA subspace model and LSR model are fused to predict the tracking result. Experimental results demonstrate that the proposed tracking method performs robustly against the target occlusion, illumination variation, motion blur, and background clutter.Finally, the thesis summarizes the main works, innovative research achievements, and the future work.