831 results about "Linear transform" patented technology

A lossless image encoding / decoding method and apparatus. The lossless color image encoding apparatus includes a motion prediction image generator estimating a motion between a previous image and a current image and outputting a corresponding prediction image, a residue generator generating a temporal residue corresponding to a difference between a prediction image generated by the motion prediction image generator and the corresponding block of the current image with respect to different components of the color image, a prediction residue generator generating prediction residues by defining a linear-transformed value of one residue among the different components of the color image output from the residue generator as a predictor and using differences between each of the residues of the other components and the predictor, and an entropy encoder receiving the residue forming the predictor from the residue generator and the prediction residues from the prediction residue generator and generating a bitstream. Encoding methods, decoding apparatuses, and decoding methods can be implemented similarly.

A method of generating matrix factors for a finite-dimensional linear transform using a computer. The linear transform is represented by data values stored in a linear transformation matrix having a nonzero determinant. In one aspect, a first LU-decomposition is applied to the linear transformation matrix. Four matrices are generated from the LU-decomposition, including a first permutation matrix, a second permutation matrix, a lower triangular matrix having a unit diagonal, and a first upper triangular matrix. Additional elements include a third matrix Â, a signed permutation matrix Π such that A=ΠÂ, a permuted linear transformation matrix A′, a second upper triangular matrix U1, wherein the second upper triangular matrix satisfies the relationship Â=U1A′. The permuted linear transformation matrix is factored into a product including a lower triangular matrix L and an upper triangular matrix U. The linear transformation matrix is expressed as a product of the matrix factors.

A computerized system and method for large scale semi-supervised learning is provided. The training set comprises a mix of labeled and unlabeled examples. Linear classifiers based on support vector machine principles are built using these examples. One embodiment uses a fast design of a linear transductive support vector machine using multiple switching. In another embodiment, mean field annealing is used to form a very effective semi-supervised support vector machine. For both these embodiments, the finite Newton method is used as the base method for achieving fast training

An image desired to be reproduced is scanned to determine its video pixel grey values. A histogram generator generates a histogram distribution representing a frequency of the grey values. The histogram distribution is analyzed to determine a minimum and maximum input grey values which define input boundaries. A segment point is computed between the input boundaries based on the histogram data. The segment point defines a plurality of input segments between the input boundaries. A dynamic output range is selected. Each input segment is mapped to an output segment based on a linear transformation for the corresponding segment. In this manner, a tone reproduction curve map having a piecewise linear transformation is automatically generated from the image histogram data.

This invention relates to a method of performing eye gaze tracking of at least one eye of a user, by determining the position of the center of the eye, said method comprising the steps of:detecting the position of at least three reflections on said eye,transforming said positions to spanning a normalized coordinate system spanning a frame of reference, wherein said transformation is performed based on a bilinear transformation or a non linear transformation e.g. a möbius transformation or a homographic transformation,detecting the position of said center of the eye relative to the position of said reflections and transforming this position to said normalized coordinate system,tracking the eye gaze by tracking the movement of said eye in said normalized coordinate system.Thereby calibration of a camera, such as knowledge of the exact position and zoom level of the camera, is avoided. Further, it is not necessary to know the position of light sources. This results in a much more flexible and user friendly system.

A method is disclosed for recovering image memory capacity, in relation to an image which has been encoded using a linear transform according to a layer progressive mode into L layers, L being an integer value greater than unity, the L layers being stored in an image memory having a limited capacity. The method comprises defining a Quality Reduction Factor (700), being a positive integer value, identifying at least one of the L layers corresponding to the Quality Reduction Factor, and discarding (702) said at least one of the L layers in progressive order in accordance with the Quality Reduction Factor, thereby recovering said memory capacity.

The invention discloses a feature-based airborne laser point cloud and image data fusion method. The method sequentially comprises the steps of conducting SIFT image stitching on image data acquired by an airborne camera; determining feature points, corresponding to image data, in point cloud data in a scanning area; establishing the corresponding relationship between the line and row numbers of a three-dimensional point cloud and line and the row numbers of pixel points in a two-dimensional stitched image plane by means of the direct linear transformation method in the photogrammetry field; finally applying RGB color information of the pixel points to every point in the point cloud so that fusion of the two kinds of data can be realized. The feature-based airborne laser point cloud and image data fusion system and method are easy to realize, high in fusion effect and fusion accuracy, capable of well achieving point cloud mapping in a small unmanned helicopter laser scanning system, high in flexibility, and adaptable to the situation that a camera and a laser scanner acquire data separately.

The invention discloses a method for multi-input multi-output transmission for precoding. The method comprises the following steps: a fundamental matrix is selected, a column vector is selected from the fundamental matrix according to the dimension of a precoding matrix under each rank, the precoding matrixes are combined into codebooks, and the codebooks are respectively stored into the transmitting end and the receiving end of a multi-input multi-output transmission system; the receiving end selects precoding matrixes from the codebooks, and feeds back the indexes thereof to the transmitting end; the transmitting end uses the precoding matrixes corresponding to the indexes when the next transmission is performed, performs linear transformation in spatial domain to signals to be transmitted and then transmits the signals. The invention also discloses a method for multi-input multi-output codebook coding for precoding. The method enables the system to meet the feedback expenses and obtains good performance gain.

A computer-implemented method is described for speaker adaptation in automatic speech recognition. Speech recognition data from a particular speaker is used for adaptation of an initial speech recognition acoustic model to produce a speaker adapted acoustic model. A speaker dependent differential acoustic model is determined that represents differences between the initial speech recognition acoustic model and the speaker adapted acoustic model. In addition, an approach is also disclosed to estimate speaker-specific feature or model transforms over multiple sessions. This is achieved by updating the previously estimated transform using only adaptation statistics of the current session.

A method is provided of transmitting signals from two or more antennas in a wireless telecommunications network, in which at least one data sequence is space-time block encoded. Before transmitting the data sequence, a linear transformation is applied to the data sequence, the linear transformation being adapted to use knowledge of correlation among the antennas to at least partially compensate the transmitted signals for said correlation. The linear transformation depends on the eigenvalues of an antenna correlation matrix. The linear transformation further depends on a ratio of symbol energy (Es) to noise variance (σ2). The method includes transmitting the encoded and transformed data sequence.

Methods and apparatus are provided for spread spectrum communication having spectrum awareness and interference mitigation based on parameterizing an input signal having multiple mixed baseband signals. The method includes non-linearly transforming the input signal, removing carrier effects, removing higher order terms by passing the input signal through a low pass filter to produce a linear combination of data symbols and DC components, solving for the DC components, separating a plurality of mixed baseband signals from the input signal, and coordinate transforming each of the separated signals from polar coordinates to a Cartesian coordinates. A receiver is provided having: a non-linear transform operator receiving a digitized input signal and adapted to produce a linear combination of data symbols, DC components, and carrier effects; a low-pass filter connected to the non-linear transform for removing higher order terms; a separator unit for separating the plurality of mixed baseband signals; and, a coordinate transform for converting the separated signals from polar coordinates to Cartesian coordinates.

An object of the present invention is to facilitate dealing with noisy speech with varying SNR and save calculation costs by generating a speech model with a single-tree-structure and using the model for speech recognition.Every piece of noise data stored in a noise database is used under every SNR condition to calculate the distance between all noise models with the SNR conditions and the noise-added speech is clustered. Based on the result of the clustering, a single-tree-structure model space into which the noise and SNR are integrated is generated (steps S1 to S5). At a noise extraction step (step S6), inputted noisy speech to be recognized is analyzed to extract a feature parameter string and the likelihoods of HMMs are compared one another to select an optimum model from the tree-structure noisy speech model space (step S7). Linear transformation is applied to the selected noisy speech model space so that the likelihood is maximized (step S8).

The invention provides a self-adaptive video image enhancement method on the basis of brightness detection. The method of the invention detects the images under different illumination conditions and enhances the low-illumination images according to the detection result. When the images are detected, average values of brightness values of all sampling points of YUV-format images in a RGB colour space are taken as the brightness value of current frame of image. For each frame of image to be processed, if the brightness value of the frame of image is less than the prearranged brightness limit value of image under normal illumination, piecewise linearity is transformed for the value of Y-component of each pixel point in the area to be processed of the frame of image so as to improve the visible effect of the low-illumination image. In order to quicken the operation speed, the brightness values of all sampling points of the image in the RGB colour space are gained and the linear transform on the Y-component of the image under low illumination is operated in a form of table lookup. The method of the invention has small calculated quantities, meets the requirement of real-time performance and can be used for the processing of video images in video communication of mobile phones and televisions.

Feature attributes are extracted from an observation space to create feature vectors for each class to be identified. A linear transformation matrix is used to reduce the dimension of the feature vectors. A numerical optimization algorithm maximizes a geometric criterion function in order to calculate the linear transformation matrix, where it exploits the geometry of the class contours of constant density. Next, a classifier based on the feature vectors in a lower dimension is generated and a class is determined for the data represented.

A system and method for semantic extraction using a neural network architecture includes indexing each word in an input sentence into a dictionary and using these indices to map each word to a d-dimensional vector (the features of which are learned). Together with this, position information for a word of interest (the word to labeled) and a verb of interest (the verb that the semantic role is being predicted for) with respect to a given word are also used. These positions are integrated by employing a linear layer that is adapted to the input sentence. Several linear transformations and squashing functions are then applied to output class probabilities for semantic role labels. All the weights for the whole architecture are trained by backpropagation.

The present invention presents a framework for separating specular and diffuse reflection components in images and videos. Each pixel of the an M-channel input image illuminated by N light sources is linearly transformed into a new color space having (M-N) channels. For an RGB image with one light source, the new color space has two color channels (U,V) that are free of specularities and a third channel (S) that contains both specular and diffuse components. When used with multiple light sources, the transformation may be used to produce a specular invariant image. A diffuse RGB image can be obtained by applying a non-linear partial differential equation to an RGB image to iteratively erode the specular component at each pixel. An optional third dimension of time may be added for processing video images. After the specular and diffuse components are separated, dichromatic editing may be used to independently process the diffuse and the specular components to add or suppress visual effects. The (U,V) channels of images can be used as input to 3-D shape estimation algorithms including shape-from-shading, photometric stereo, binocular and multinocular stereopsis, and structure-from-motion.

A method is disclosed herein that derives a post-alignment transform that is applied to the input source after the application of an alignment transform to ensure that when each source illuminates the screen, the display is aligned and retains the correct displayed aspect ratio. The transform takes into account the alignment transform and the input source aspect, and then performs a secondary, post-processing transformation that corrects the distortion induced by the alignment transform. This aspect-preserving function is a linear transform that maps pixels in the aligned space to the aligned and aspect-correct space and can be applied to the input signal after the first alignment transform, or, in the case where the alignment transform is also linear, transforms can be directly combined to derive a compound multi-projector, aspect preserving transform. In this way, the input signal to the projectors will be aligned as well as retain proper aspect.

A measured phasor representing a sinusoidal power signal waveform is produced by orthogonal FIR filters, one of which produces the real part of the phasor and the other of which produces the imaginary part. The frequency difference between the rated frequency of the signal waveform in the power system and the actual frequency of the waveform at any selected point in time is determined and the elements (A, B, C, D) of a compensation matrix is determined. The compensation matrix is then used to perform a linear transformation to produce an exact phasor for the waveform.

The invention discloses a multilayer neural network language model training method and device based on knowledge distillation. The method comprises the steps that firstly, a BERT language model and amulti-layer BILSTM model are constructed to serve as a teacher model and a student model, the constructed BERT language model comprises six layers of transformers, and the multi-layer BILSTM model comprises three layers of BILSTM networks; then, after the text corpus set is preprocessed, the BERT language model is trained to obtain a trained teacher model; and the preprocessed text corpus set is input into a multilayer BILSTM model to train a student model based on a knowledge distillation technology, and different spatial representations are calculated through linear transformation when an embedding layer, a hiding layer and an output layer in a teacher model are learned. Based on the trained student model, the text can be subjected to vector conversion, and then a downstream network is trained to better classify the text. According to the method, the text pre-training efficiency and the accuracy of the text classification task can be effectively improved.

A system and methods are provided for extending the frequency bandwidth of a harmonic signal. Harmonic content of a band-limited signal is extended to frequencies outside the signal's passband by performing a non-linear transformation on the complex spectrum of the band-limited signal in the frequency domain. The non-linear transformation may be accomplishes by a linear convolution of the complex spectrum with itself. A system for extending the frequency bandwidth of a harmonic signal includes a signal processor with a forward transform module for transforming a time domain signal into the frequency domain, a non-linear transform module for performing the non-linear transformation on the complex spectrum of the harmonic signal, and a reverse transform module for transforming the extended spectrum of the harmonic signal back into the time domain.

The invention involves a calibration device for camera calibration, which consists of two vertical grid boards that have distribution grid and the number of signs marked with a unique code point. The invention also provides a method of the camera calibration, including the following steps: (a) Acquisition of images of the calibration device; (2) corresponding coordinates of the signs point to the coordinates of object point; (3) projection the plane grid to get the initial point coordinates of grid point; (4) the use of line extraction intersection to get the precise coordinates of points; (5) The use of 3-D direct linear transformation for the initial value of calibration parameters; (6) iterative solution with beam adjustment to get the camera's internal and external orientation elements. After shooting a picture of calibration device in the invention, the follow-up calibration process can be done automatically by computer. The advantage of the invention is automatic and quickness.

An optical touch-sensitive device has the capability to determine touch locations of multiple simultaneous touch events. The touch events disturb optical beams propagating across the touch sensitive surface. With multi-touch events, a single beam can be disturbed by more than one touch event. In one aspect, a non-linear transform is applied to measurements of the optical beams in order to linearize the effects of multiple touch events on a single optical beam. In another aspect, the effect of known touch events (i.e., reference touches) is modeled in advance, and then unknown touch events are determined with respect to the reference touches.

The invention discloses a plaintext or ciphertext selection based side channel power analysis attack method on round function output of the SM4 cipher algorithm. The method includes the steps of S1, selecting plaintext or ciphertext to input X<i+0>, X<i+1>, X<i+2> and X<i+3> on the condition of allowing the exclusive OR result of the X<i+1>, X<i+2> and X<i+3> to be a constant value and guaranteeing randomness of the X<i+0>, utilizing side channel power attack processes to attack the output C of linear transform L of each round of the first four round functions, and deducing inversely to acquire the round key rk of the first four round functions in encryption or decryption according to the output C; S2, according to the round keys rk<0>, rk<1>, rk<2> and rk<3> of the first four found functions, inversely calculating the initial key by a key expansion algorithm. The method can realize power analysis attack by means of multiple attacks and can attack with selection of proper-length bits according to actual computing capacity, thus flexibility, effectiveness and success rate of analysis are improved.

The invention relates to an encryption method for network and information security; the adopted packet length and key length are 128 bits, and a 128-bit system parameter is added; encryption algorithm comprises a plurality of rounds of round transform iteration consisting of mixed layers and diffusion layers, wherein the last round is incomplete round transformation, and the incomplete round transformation consists of the mixed layer; the mixed layer comprises key plus, S box portfolio transform and system parameter minus; the S box portfolio transform comprises 8-input 8-output reversible S box transform and inverse S box transform in the same number which are output by parity exchange; the diffusion layer can construct linear transformation with a reversible re-model polynomial matrix; an encryption round key is generated by an encryption key through encryption round transformation; and decryption algorithm is the inverse transformation of encryption algorithm. The encryption methodfor network and information security has the advantages of high diffusion speed, good security strength, hardware resource saving, and very high speed when being realized by hardware and on a software platform.

A privacy-preserving scalar product calculation system is provided. A first unit linearly transforms an n-dimensional vector Va into an n-dimensional vector based on a scalar value based on a random number Wi and a random number Rj to calculate a remainder by dividing each element of the linearly transformed n-dimensional vector by a random number Mi, and transmits an n-dimensional converted vector X including each of the remainders as its element to the second unit, the second unit calculates an inner product value Z based on the received n-dimensional converted vector X and an n-dimensional vector Vb, and transmits the inner product value Z to the first unit, and the first unit further calculates, based on a reciprocal of the scalar value and the receive inner product value, a scalar value and which calculates a remainder by dividing the scalar value by the random number Mi.

A system and method is described for rendering a warped brush stroke using a bitmap brush image, the brush stroke being along a arbitrarily curved guideline. The described system and method generate a piecewise linear approximation to the guideline followed by generating polygons with the aid of the linear segments such that the generated polygons are convex and contiguous linear segments result in contiguous polygons. A mapping is identified between segments of the bitmap brush and the polygons such that the corners or the boundaries of the segments of a segment map to the corners or boundaries of a corresponding polygon. The segment of the bitmap brush is mapped into the corresponding polygon using transformations that do not require visiting a pixel in the rendered warped brush stroke more than once. Examples of such transformations include bilinear transformations and texture mapping in combination with tiling.

A linear transform such as a Fast Fourier Transform (FFT) is performed on an input data set having a number of points using one or more arrays of concurrent threads that are capable of sharing data with each other. Each thread of one thread array reads two or more of the points, performs an appropriate “butterfly” calculation to generate two or more new points, then stores the new points in a memory location that is accessible to other threads of the array. Each thread determines which points it is to read based at least in part on a unique thread identifier assigned thereto. Multiple transform stages can be handled by a single thread array, or different levels can be handled by different thread arrays.

A method is provided of transmitting signals from two or more antennas in a wireless telecommunications network, in which at least one data sequence is space-time block encoded. Before transmitting the data sequence, a linear transformation is applied to the data sequence, the linear transformation being adapted to use knowledge of correlation among the antennas to at least partially compensate the transmitted signals for said correlation. The linear transformation depends on the eigenvalues of an antenna correlation matrix. The linear transformation further depends on a ratio of symbol energy (ES) to noise variance (σ2). The method includes transmitting the encoded and transformed data sequence.

The invention discloses an extraction method for brain-computer interface system imagination action EEG signal features, in particular to an EEG feature extraction method based on a wavelet transform and a BP neural network. The invention takes the energy change caused by imagination action thinking to be a feature distinguishing the imagination movements of a left hand and a right hand, respectively calculates the point-to-point average power of the entire samplings of the EEG signal obtained from C3 and C4 channels by the left hand and the right hand through the imagination (thereinafter called as C3 and C4 of the left hand and the right hand) within 0 to 9s according to the average power formula. A time window is arranged, a discrete dyadic wavelet transform is made to the data of a section provided with the window, an approximation signal a6 on a sixth size is selected to be taken as a signal feature; a BP neural network is used as a classifier to classify. The method of the invention adopting the wavelet transform and the BP neural network to extract the potential of the imagination movement helps to improve the signal / noise ratio and the identification correction rate of the potential of the imagination action; in addition, the wavelet transform is a linear transform, has a quick calculation speed, and is suitable for on-line analysis.