Search

This paper suggests a scheme of image denoising based on two-dimensional discrete wavelet transform. The denoising algorithm is described with some operators. By thresholding the wavelet transform coefficients of noisy images, the original image can be reconstructed correctly. Different threshold selections and thresholding methods are discussed. A new robust local threshold scheme is proposed. Quantifying the performance of image denoising schemes by using the mean square error, the performance of the robust local threshold scheme is demonstrated and is compared with the universal threshold scheme. The experiment shows that image denoising using the robust local threshold performs better than that using the universal threshold.

In this paper, we proposed a novel scheme to recognize 3D face based on partial differential equations (PDE). We proposed a novel distance between two boundary curves using the difference of the corresponding Fourier series coefficients. The distance of each adjacent boundary curves is taken as the feature of 3D face for recognition. Moreover, resemblance distance algorithm is applied for recognition. Experiment results and analysis show that the proposed scheme has a high identification rate in neutral 3D face database.

Dimensionality reduction algorithms, which aim to select a small set of efficient and discriminant features. In this paper, a new manifold learning algorithm, called Uncorrelated Locality Information Projection (ULIP), to identify the underlying manifold structure of a data set. ULIP considers both the between-class scatter and the within-class scatter in the processing of manifold learning. Equivalently, the goal of ULIP is to preserve the intrinsic graph characterizes the interclass compactness and connects each data point with its neighboring points of the same class. Different from Principal component analysis(PCA)that aims to find a linear mapping which preserves total variance by maximizing the trace of feature variance and the optimal mapping is the leading eigenvectors of the total variance matrix associated with the leading eigenvalues, While locality preserving projections(LPP) that is in favor of preserving the local structure of the data set. we choose proper dimension of subspace that detects the intrinsic manifold structure for classification tasks. Extensive experiments on face recognition demonstrate that the new feature extractors are effective, stable and efficient.

In order to avoid pixel expansion and contrast of the gathered secret image compromise, an easier visual cryptographic scheme, which was called visual cryptographic of interleaving with sub-block coding method, was presented. Instead of using m subpixels to represent one pixel, according to a key matrix, the secret image was divided into 2 shadow images, in which there were some background pixels useless. In shadow images, each pixel copied from the secret image maintained the exact value and position in the original secret image, therefore the secret image could be superposed, with interleaving, from the 2 shadow images exactly. Then sub-block coding method was adopted to convert each shadow image to numbers matrix, that is, every block met an unique number. After that, according to a secret key, these two number matrices were converted into a series of number. Last the series was encrypted to transport or do some other operations. This scheme was compared to the (k, n) Visual Secret Sharing scheme in Section 4 and it had been shown that there did not exist pixel expansion and contrast compromise in this new scheme, thus a better effect was presented.

The objective of this work is the detection of object classes. An improved method is used for object detection and segmentation in real-world multiple-object scenes. It has two stages. In the first stage this method develops a novel technique to extract class-discriminative boundary fragments and the texture features near the boundary, and then boosting is used to select discriminative boundary fragments (weak detectors) to form a strong ?boundary-fragment-model? detector. An appearance model is built with those entire detectors and the texture features. In the second stage, the boundary fragment and the texture features and used to complete detection. To the end, a new fast cluster algorithm is used to deal with the centroid image. The generative aspect of the model is used to determine an approximate segmentation. In addition, we present an extensive evaluation of our method on a standard dataset and compare its performance to existing methods from the literature. As is shown in the experiment, our method outperforms previously published methods with the overlap part of the object in multiple-object scene.

Temporal synchronization is one of the challenges for blind watermark detection. If the detector?s input is watermarked but synchronization fails, then the embedded watermark will not be detected. In this paper, a novel video watermarking algorithm is proposed based on data statistics classification of motion vectors. Through analyzing data statistics of motion vectors? magnitude, we classify the motion vectors into several groups, and then according to a certain rule, we embedded one bit of watermark sequence into motion vectors in one group. Thus, watermark is associated with motion vectors? statistical characteristics, and those motion vectors carrying the identical watermark bits are uniformly distributed to the whole video sequence. It avoids the sensitivity to watermark synchronization comparing to traditional algorithm. The experimental results demonstrate that this scheme is robust to common kinds of video attack, including frame dropping, frame inserting etc. Moreover, the scheme has the property of perceptual transparency and random detection.

In this paper, we propose a new manifold learning algorithm, called Uncorrelated Locality Preserving Projections, to identify the underlying manifold structure of a data set. ULPP tries to find the subspace that best discriminates different face classes by maximizing the between-class distance, while minimizing the within-class distance. Different from Principal component analysis(PCA)that aims to find a linear mapping which preserves total variance by maximizing the trace of feature variance and locality preserving projections(LPP) that is in favor of preserving the neighborhood structure of the data set. We choose proper dimension of subspace that detects the intrinsic manifold structure for classification tasks. Experiments comparing the proposed algorithm with some other popular algorithms on the JAFFE, AT&T, and Yale databases show that our algorithm consistently outperforms others.

Dimensionality reduction algorithms, which aim to select a small set of efficient and discriminant features. In this paper, a new manifold learning algorithm, called uncorrelated maximum locality preserving projections(UMLPP), to identify the underlying manifold structure of a data set. UMLPP considers both the between-class scatter and the within-class scatter in the processing of manifold learning. Equivalently, the goal of UMLPP is to preserve the intrinsic graph characterizes the interclass compactness and connects each data point with its neighboring points of the same class. Different from principal component analysis (PCA) that aims to find a linear mapping which preserves total variance by maximizing the trace of feature variance, While locality preserving projections (LPP) that is in favor of preserving the local structure of the data set. We choose proper dimension of subspace that detects the intrinsic manifold structure for classification tasks. Extensive experiments on face recognition demonstrate that the new feature extractors are effective, stable and efficient.

The objective of this work is the detection of object classes. An improved method is used for object detection and segmentation in real-world multiple-object scenes. It has two stages. In the first stage this method develops a novel technique to extract class-discriminative boundary fragments, and then boosting is used to select discriminative boundary fragments (weak detectors) toform a strong "boundary-fragment-model" (BFM) detector. A boundary fragment dictionary is built with those entire detectors. In the second stage, after edge detection, length filter is used to improve the match degree. To the end, a new fast cluster algorithm is used to deal with the centroid image. The generative aspect of the model is used to determine an approximate segmentation. In addition, we present an extensive evaluation of our method on a standard dataset and compare its performance to existing methods from the literature. As is shown in the experiment, our method outperforms previously published methods with the overlap part of the object in multiple-object scene.

In order to merge information from multi-sensor adequately, this paper analyzed the typical remote sensing image fusion methods. It proposed a new fusion method based on window local deviation using WPT (wavelet packet transform). WPT decomposed an image into low frequency hand and high frequency hand on different levels; local window and local sub-window statistic in the wavelet domain were used to further classify the wavelet coefficients into the edge and non-edge coefficients. In the fusion processing, the fused approximate coefficients were obtained by average method and the fused detailed coefficients were obtained by different area-based feature selection method and the greatest multi-window local deviation of the corresponding enter image wavelet packet coefficients. Experimental results show that the proposed method can retain the subtle detail of the image and can achieve an excellent effect. The new image fusion method provides a performance better than other methods.

One of the challenges in blind watermark detection is synchronization. Synchronization is the process of identifying the coordinates of an embedded watermark and is crucial in successful watermark detection. If the detectorpsilas input is watermarked but synchronization fails, then the embedded watermark will not be detected. In this paper, novel video watermarking algorithm is proposed basing on date statistics of motion vectors. According to the amplitude value of motion vectors, we divide the motion vectors into groups, and then choose some groups for embedding watermark. One bit of watermark sequence is embedded into one group. The embedded watermark information relates closely with the amplitude value of motion vectors. Therefore, our watermark frame avoids the sensitivity to watermark synchronization compared to traditional algorithm. The experimental results demonstrate this scheme is robust to common kinds of video attack, including frame losing, frame inserting etc. Moreover in the scheme, there is the property of perceptual transparency and real-time.

In pattern recognition, feature extraction techniques are widely employed to reduce the dimensionality of data. In this paper, a new manifold learning algorithm, called Enhanced Locality Preserving Projections, to identify the underlying manifold structure of a data set. ELPP considers both the between-class scatter and the within-class scatter in the processing of manifold learning. Equivalently, the goal of ELPP is to preserve the within-class geometric structure, while maximizing the between-class distance. Different from Principal component analysis (PCA) that aims to find a linear mapping which preserves total variance by maximizing the trace of feature variance and the optimal mapping is the leading eigenvectors of the total variance matrix associated with the leading eigenvalues, While locality preserving projections(LPP)that is in favor of preserving the local structure of the data set. We choose proper dimension of subspace that detects the intrinsic manifold structure for classification tasks. Experimental results on UMIST face database showed ELPP can represent class separability and clustering performance better than LDA and MMC. Extensive experiments on face recognition show the effectiveness of the proposed ELPP method.