The traditional algorithms of image super-resolution reconstruction are not effective enough to be used in reconstructing high-frequency information of an image. In order to improve the quality of image reconstruction and restore more high-frequency information, the residual dictionary is introduced which can capture the high-frequency information of images such as the edges, angles and corners. The common dictionary is generated by training and learning pairs of low-resolution and high-resolution images. The dictionary combined by common dictionary and residual dictionary is obtained in which more high-frequency information of the images can be restored while the spatial structure of images can be preserved well. The processing of training and testing dictionary is conducted by Support Vector Regression (SVR). Compared with other algorithms in experiments, the proposed method improves its PSNR and SSIM by 3% ~ 4% and 2% ~ 3% on some different images respectively.
In this paper, we propose a algorithm to tracking target using point feature. The point feature is extracted from the pixels in the first frame and used to label the pixels in the next frame as belonging to either target or background. The positive and negative samples are extracted from the pixels of target and surrounding background, and used to train several weak classifiers, which combine to build a strong classifier using AdaBoost algorithm. The negative samples are given the greater weights than positive samples, which is to avoid that a large number of pixels in background are labeled incorrectly. To efficiently learn a large number of samples, the adopted weak classifier is a linear perceptron model, which is trained and updated using stochastic gradient descent. Only the dot-product between matrices and the sum of matrix elements need to be calculated. To distinguish the similar targets, the histogram-based mean shift algorithm is applied to eliminate those wrong image patches. The histogram of target will be updated over the time. The experiment results show that the proposed algorithm can estimate scale better when scale change, posture change and occlusion occurs.
Due to the limitations of image capture device and imaging environments in traditional imaging process, high-resolution (HR) images are difficult to be obtained. The method of digital image processing can be used in image super-resolution with one or an image sequence in original conditions to reconstruct HR images which over the range of imaging system. Traditional learning-based super-resolution algorithm need to run through the sample library with a high computing complexity, and a high recognition rate in the scene with small shifts. This dissertation is mainly about color image SR and parallel implementation of the SR algorithm. An algorithm based on SVM classified learning is proposed in this paper.