Abstract:Traditional NMF method does not fully utilize the internal similarity among original data, thus the performance of dimensionality reduction is limited. To this end, a new nonnegative matrix facto rization algorithm restrained by the regularization of potential information is proposed. Firstly, the potential information is mined via the iterative nearest neighbor. Then the potential information is utilized to construct similarity graph of data set. Finally, the similarity graph is incorporated as a regularization term to preserve the relationship between original data in the decomposition process of nonnegative matrix. The regularization term keeps the similarity between the original data in the process of dimensionality reduction, which can improve the discriminant ability of nonnegative matrix factorization algorithm. Thorough experiments on standard image databases show the superior performance of the proposed method. Traditional NMF method does not fully utilize the internal similarity among original data, thus the performance of dimensionality reduction is limited. To this end, a new nonnegative matrix facto rization algorithm restrained by the regularization of potential information is proposed. Firstly, the potential information is mined via the iterative nearest neighbor. Then the potential information is utilized to construct similarity graph of data set. Finally, the similarity graph is incorporated as a regularization term to preserve the relationship between original data in the decomposition process of nonnegative matrix. The regularization term keeps the similarity between the original data in the process of dimensionality reduction, which can improve the discriminant ability of nonnegative matrix factorization algorithm. Thorough experiments on standard image databases show the superior performance of the proposed method. Traditional NMF method does not fully utilize the internal similarity among original data, thus the performance of dimensionality reduction is limited. To this end, a new nonnegative matrix facto rization algorithm restrained by the regularization of potential information is proposed. Firstly, the potential information is mined via the iterative nearest neighbor. Then the potential information is utilized to construct similarity graph of data set. Finally, the similarity graph is incorporated as a regularization term to preserve the relationship between original data in the decomposition process of nonnegative matrix. The regularization term keeps the similarity between the original data in the process of dimensionality reduction, which can improve the discriminant ability of nonnegative matrix factorization algorithm. Thorough experiments on standard image databases show the superior performance of the proposed method.

Abstract:Word embedding refers to a machine learning technology which maps search of word lying in high-dimensional discrete space (with dimension to be the number of all words) to a real number vector in low-dimensional continuous space. Word embedding provides better semantic word representations, and　thus greatly benefits text processing tasks. Meanwhile, huge amount of unlabeled text data, together with the development of advanced machine learning techniques such as deep learning, make it possible to effectively obtain high quality word embeddings. Besides, the definition and practical value of word embedding are given, and some classical methods are also reviewed to obtain word embedding, including neural network based methods, restricted Boltzmann machine based methods, and methods based on factorization of context co-occurrence matrix. For each model, its mathematical definition, physical meaning are introduced in detail, as well as training procedure. In addition, all these methods are compared in the aforementioned three aspects.

Abstract:TV-Wavelet-L1（TVWL1）model which consists of total-variation (TV) and wavelet regularization has great capability in image reconstruction. However, traditional algorithms solving the TVWL1 model for image reconstruction ignore the way of synthesis/analysis sparse representation. A new image reconstruction algorithm is thus proposed to solve TVWL1, where the original signal reconstruction problem is decomposed into multiple much simpler sub-problems which can be solved alternately. In addition, the analysis sparse representation is considered in a sub-problem. Experimental results demonstrate that the proposed algorithm can obviously improve both objective and subjective qualities of reconstruction images compared with the existing algorithms.

Abstract:In this work, a novel video compression framework with super-resolution technique is proposed. The input image is first down sampled by down sampling factor 2. Then the down sampled image is coded by JPEG standard. A novel hybrid super-resolution (SR) method is applied to decoded down sampled image. Meanwhile, feedback is designed to further improve the quality of final decoded video. Specifically, by original image subtracts super-resolution image is residual assistance image at encoder side. Then, this residual assistance image can be compensated for the loss of high-frequency details in SR process at decoder side. Moreover, only one quantization parameter (QP) to control the quality of coding image is needed for the whole framework. Evaluations have been made in comparison with JPEG standard coding scheme. Experimental results show that proposed coding framework achieves significant bitrate saving and compression ratio increase at similar objective quality levels.

Abstract:Some studies have shown that optimizing the projection matrix can improve the reconstruction of compressed sensing and the sparsity range of signal adaption. This method uses iterative updated Gram matrix to maximum the optimization of compressed sensing(CS) projection matrix. It is a new method for enhancing the CS performance, which is different from previous design problems of projection matrix. Here,it analyzes, summarizes and compares the structure of those existing optimization methods of projection matrix, the application characteristics as well as existing problems, and concludes with the discussion of its possible direction of future development. The experimental results are used to verify the analysis of the conclusions.

Abstract:As an important automatic identification technology, the two-dimensional bar code has wide range of prospects for commercial applications. Effective solution of the two-dimensional bar code image blur is the key to its wide use. In order to recognize the blurring type, an algorithm based on invariant moment theory is proposed. The frequency spectrum in different types of degradation for blurred bar code images is analyzed to find their differences. After edge detection, denoise processing and binarization , the invariant moment feature is extracted to recognize the blurring type. Experimental results show that the proposed algorithm can obtain a high recognition rate.

Abstract:It is difficult to detect the target of aircrafts in high resolution remote sensing images. So a bag of visual words model is proposed. In order to abtain the most discr imi native visual words and make visual codebook effective and compact, the visual w ords that are irrelevant, weakly irrelevant and redundant must be removed from t he visual codebook. The method integrating relevance analysis wit h redundancy analysis is used to prune out those useless visual words. Finally, t he most important visual words are chosen to describe aircrafts in high resoluti on remote sensing images, which helps reducing the computation in the following test stage and also improving the efficiency.

Abstract:Ultrasound elastography has been well applied to medical detection as a tool for diagnosis aid. However, in conventional ultrasound elastograms, there are patterned artifacts from correlated errors. And random fluctuations of the signal amplitude result in a match only occurrs at certain regions from pre-and post-compression window pairs. The artifacts should be prevented from diagnosis. A spatial displacement compounding method is proposed to correct the errors artifact reduction. It is proved that the method is in correct principle. The presented results from a commercial elastic phantom show some improvement in SNRe and CNRe.

Abstract:According to the phenomena that the calculation of prior probability in text classification is time-consuming and has little effect on the classification result，and the accuracy loss of posterior probability affects the accuracy of classification, the classical naive Bayes algorithm is improved and a new text classification algorithm is proposed which restrains the effect of prior probability and amplifies the effect of posterior probability. In the new algorithm, the calculation of prior probability is removed and an amplification factor is added to the calculation of posterior probability. The experiments prove that removing the calculation of prior probability in text classification can accelerate the classification speed and has little effect on the classification accuracy, and adding an amplification factor in the calculation of posterior probability can reduce the effect of error propagation and improve the classification accuracy.

Abstract:The denoising and beautification for foreground image is an important segment in the fields of pattern recognition and computer vision. Given the distortion of foreground in geometry and dimension caused by morphological post-processing method, a new denoising method is presented by using the layered filter about independent separation block whose pixels’ spatial distribution and value are different for the actual foreground and noise in statistics to filter noise. The experimental results show that the proposed method can effectively wipe off the noise on the foreground image, and perform better as compared to the morphological image processing, and has a lower time complexity.

Abstract:Gene expression data has a high application value for understanding the pathogenesis, disease diagnosis and gene-level drug development. However, the microarray data usually contains thousands of genes with a small number of samples, which causes serious curse of dimensionality and deteriorates the diagnosis accuracy. Moreover, it gives raise to difficulty to a lot of classifiers, and cuts down the cost of medical diagnosis. A new gene selection method is proposed, which is based on clustering and particle swarm optimization (PSO). Firstly, partition the genes using clustering algorithm and the useful are clusters selected for classification. Then the wrapper selection method based on particle swarm optimization(PSO） and extreme learning machine(ELM） is used to select the compact gene subset with high classification accuracy from the genes selected before. This method take advantages of clustering and PSO algorithm, and it can perform better in classification than other classical methods.

Abstract:Diffusion weighted image (DWI) is multi-boundary in nature, and it is particularly important to get accurate boundary signals of DWI . A new method combing bidimensional empirical mode decomposition (BEMD) with modified adaptive Wiener filter is proposed. Through BEMD method the degraded image is decomposed into a detail part and a residual part. The detail part of the image contains the boundary signal and the noise of the degraded image, and the residual part describes the image tendency. Thus, the DWI data is decomposed by BEMD method at first, and then, the modified adaptive Wiener filter is applied to remove the noise in the detail part of DWI and the residual part is thus handled. Finally, the denoised detail image and its residual are combined to form the denoised DWI. The method is performed on real DWI data. Experiment results positively show that the proposed method removes noise effectively and keeps the boundary of DWI successfully.

Abstract:Since traditional shadow detection methods based on paired-regions are prone to over-segmentation in complex texture regions, they always have high computational complexity which affects the detection result to a certain extent. Therefore, an improved shadow detection algorithm is presented, using clustering method to merge divided areas and decrease over-segmentation. Besides, support vector machine (SVM) is established to classify the features of paired regions, so both algorithm efficiency and shadow detection effect are improved. Simulation results indicate that the running time of the proposed algorithm is remarkably reduced compared with the traditional one. Moreover, the shadow detection effect is more accurate than that of the original method for complicated texture images.

Abstract:Region of interest (ROI) encoding is an important feature of SVAC. The efficiency of image processing can be increased and the bit rate reduces greatly because of ROI, but the bit rate will fluctuate along with complexity of encoding frames. Therefore, availability of communication channel is inefficient. There has been no rate control algorithm recommended by SVAC so far. A rate control algorithm is thus provided for ROI based on SVAC. The bit rate of ROI and back region are allocated by calculating the complexity of image, and the bit rate of ROI will be allocated preferentially. Meanwhile, a virtual buffer is designed. The occupancy of virtual buffer is calculated to adjust the quantization parameter (QP) of ROI and back region of every encoding frame in time. Experiment results show that bit rate is stable in communication channel whose availability is improved. Besides the image quality of ROI is smooth and excellent.

Abstract:Taking the multiple features of remote sensing images into consideration, a new approach is presented to the classify remote sensing images based on the fusion of multiple features. The bag of visual words (BOVW) representation is firstly improved. Then, the BOVW feature, color histogram and Gabor texture feature are extracted from the images respectively. The classification is performed by the support vector machine classifier, and the final output is obtained through adaptively fusing the results by multiple features. The proposed method has been evaluated on a large publicly available remote sensing image dataset with 2 100 images. The experimental results have witnessed that the overall classification accuracy is boosted by 10% in comparison with the method based on the single feature which owns the highest accuracy. Comprehensive experimental results indicate that the proposed approach is effective and suitable for high-resolution remote sensing image classification.

Abstract:In digital image forensics, most of the works perform resampling detection by detecting periodicity introduced by the interpolation process. The detection of periodicity is usually done in frequency domain where the resampling rate can be determined simultaneously. But due to frequency aliasing, the resampling rate cannot be totally determined. To solve this problem, an algorithm is proposed which can determine the resampling rate in spatial domain, based on the observation that the redundancy of the image rows (or columns) vary periodically and can be used to determine the resampling rate. As experiments show, all resampling rates can be correctly detected in uncompressed images, and in compressed images the proposed algorithm shows more robustness compared to the prior algorithm.

Abstract:To deal with the disadvantage that classical evidence theory cannot combine highly conflicting evidence, a new combination approach is presented based on distance measurement. The evidence is regarded as space vector and distance measurement space is defined, in which the distance between evidence is calculated. Then the consistency measurement is established using distance measurement matrix by which the support degree is gained on the frame of discernment. Finally, the evidence consistency coefficient obtained from normalizing support degree is regarded as the weight to distribute conflicting probability, and improved evidence combination formula is proposed. Numerical examples prove that the improved formula can combine the conflicting evidence and the non-conflicting evidence. The contrast to other methods testifies the validity of improved formula.

Abstract:Video stitching is one important branches of computer graphics and computer vision, and it is rooted on the development of static image mosaic technology. However, due to the complexity of the video information, video stitching is different from image mosaic. Aiming at the actual demanding of real-time stitching, a video stitching algorithm of fixed cameras based on control images is proposed. Firstly, the control images are captured in order to calibrate the cameras to get the internal reference and the coordinates of optical center. An improved distortion correction algorithm has been applied to deal with the distortion of the cameras so as to control the distortion of the captured control images. The scale invariant feature transform(SIFT) features of the control images are extracted and processed through a rough matching. Then random sample consensus(RANSAC) algorithm is adopted to reduce the number of error-matching and to fit the image homograhy. The method of lookup table is applied to project each image captured from the cameras onto the panorama image. After gain compensation, the overlapped portion is blended by multi-band technique. Finally, the real-time video stitching is implemented, which can reach the final speed of 25 frames per second.

Abstract:In view of the noise problems when detecting the edge of lung images, three points are improved based on the mathematical morphology edge detection. Firstly, connected with three fundamental select principles for the structuring element, i.e. the similarity of shape, the covering of size and the composition of the different structuring elements, it particularly chooses omnidirectional structures and multi-scale structures suitable for the lung images. Secondly, it improves common morphological edge detection operator, and combines omnidirectional structures and multi-scale structures to obtain a new compound morphology edge detection operator which is suitable for edge detection of the lung images. Thirdly, it adds peak signal-to-noise ratio (PSNR) into weight calculation method and improves the method for calculating weights. Finally, it detects the edge of lung noise images with PSNR of 50.684 9 dB through the simulation. Compared with the general algorithm, The results show that the improved algorithm can improve PSNR and mean square error (MSE) substantially and can detect more and better de-noising lung image edge. Applied to other images or different noise images, the proposed algorithm can detect sharper edges of the image, indicating that the algorithm has good robustness.

Abstract:Genome-wide nucleosome prediction has been an important research area in genetics so far. However, most existing nucleosome prediction algorithms are based on the statistical features of nucleosome-bound DNA sequences, usually resulting in low accuracies. Basides our statistical studies in linker DNA sequences, each of which connects with two nucleosome-bound DNA sequences, show that linker DNA sequences have some specific statistical properties.Concerning the fact, improvement of the Segal model is presented, where two score functions are constructed, based on the dinucleotide position frequencies of the nucleosome-bound and linker DNA sequences respectively. Nucleosome positions are predicted according to the difference between the above two score functions. Experimental results on the yeast’s chromatin demonstrate that the improved algorithm can significantly increase the accuracy in positioning nucleosomes.

Abstract:Fuzzy support vector machine (FSVM) is used on road roughness recognition. The general SVM is particularly sensitive to the noise points and outliers in the samples, so a method is proposed, in which the distance from sample to the center of class is taken as the fuzzy membership of the sample and the parameters of FSVM are optimized by improved particle swarm optimization (PSO) algorithm. After training and testing the experimental data, the highest average recognition rate increases to 77.5%, which is higher than 72.5% that of the method with the general support vector machine. Data processing indicates that FSVM strengthens the influence of effective samples on classification and weaken influence of noise points and outliers. Furthermore, the recognition rate of road roughness has been improved.

Abstract:For airport noise prediction, aiming at the high cost and large error of contour drawing, as well as the lack of guidance standard in regression method based on SVM classification, it presents a method of cluster regression based on support vector machine (SVM). Cluster regression in prediction of airport noise, using k-means algorithm, takes advantage of the characteristics of clustering. It firstly limits the sample within the same class, and then performs regression in the similar class. Experimental results on housing data set and Laser generated data set show that the fitted values of the cluster regression method are more accurate than the direct regression method. Applied the method to measured data of an airport in Beijing, and compared it with other prediction models, the accuracy of cluster regression is superior to that of other prediction methods.