27 results about "Hadamard transform" patented technology

The invention belongs to the technical field of large model quantification, and discloses a large model quantification method based on non-uniform grouping Adamage transform and activation distribution self-adaption, which comprises the following steps: firstly, eliminating abnormal values in an activation matrix through non-uniform grouping Adamage transform, and solving the problem of abnormal distribution caused by uniform grouping; secondly, constructing an equivalent transformation error elimination algorithm, and eliminating a weight quantization error by adopting a GPTQ algorithm; and finally, distributing an optimal quantization data type (Int4 or Nint4) for each activation matrix through an activation distribution adaptive algorithm so as to minimize a quantization error. According to the method, the deployment efficiency of online rotation is improved through non-uniform grouping Adamage transformation, meanwhile, the quantization error is remarkably reduced in combination with self-adaptive data type selection, and the model quantization precision and the hardware deployment efficiency are improved. The method is suitable for compression and optimization of a large language model, and has a wide application prospect.

This invention discloses a deep learning-based Hadamard transform-based anti-screen-capture watermarking method, belonging to the field of watermarking technology. This invention combines a convolutional neural network (CNN) and residual blocks to achieve an end-to-end process of watermark embedding and extraction within the Hadamard domain. By simulating screen-capture attacks and incorporating this method between the embedding and extraction layers, the network ensures robust watermark embedding. This method can efficiently extract watermark information from photos taken by surreptitiously, protecting the copyright of digital media. The use of the transform domain for watermark embedding allows the watermark to spread over a wider area of ​​the image, significantly improving the robustness of the watermarking algorithm. Comparison with related technologies demonstrates the superiority of this invention in terms of imperceptibility and robustness.

The invention relates to the technical field of biological feature recognition, in particular to a face template protection method and system based on Hadamard transform and improved MinHash, a registered user face template is generated through the face template protection method, and the registered user face template is stored in a protected face template database. According to the face template protection method, a face template is generated through face image feature extraction, local Hadamard transformation and an improved MinHash algorithm; the method comprises the following steps: acquiring a face image of a to-be-authenticated user, generating a target face template of the to-be-authenticated user through a face template protection method, and performing similarity matching on the target face template and a registered user face template in a protected face template database to judge whether the face image of the to-be-authenticated user is a registered user face image or not. According to the method, two-layer protection can be realized on the face template in face recognition, the safety is relatively high, the recognition accuracy and generation efficiency of the face template are effectively improved, and the method has a relatively good application prospect in the field of face recognition identity authentication and authorization.

The invention discloses an optical computing chip with a vertical micromirror structure and a wafer-level intelligent manufacturing method thereof, the optical computing chip comprises a plurality of optical modulators and a plurality of quantum logic gates, and interference optical paths in the optical modulators are set based on a vertical micromirror, so that optical modulation is realized; the quantum logic gate constructs a stable light path through a reflecting mirror arranged vertical to the micro-mirror, interference is realized by using a beam splitter, and Hadamard transform is executed by a wave plate at an inlet or an outlet of a target light path, so that light calculation is realized; a deep groove beam splitter or a deep wall beam splitter is adopted as a beam splitter, a silicon wall in the deep wall beam splitter is arranged to be perpendicular to two side walls of a substrate to form a preset included angle, and the deep groove beam splitter is manufactured by the following steps: photoetching a pattern of a deep groove on the surface of a monocrystalline silicon wafer; etching a deep groove with a preset angle by using a deep etching method, and smoothing the side wall; depositing an optical thin film in the deep groove; photoetching and deeply etching the shape of the deep groove beam splitter, wherein the side wall of the deep groove beam splitter is smooth; and growing an optical antireflection film on the vertical side wall.

A quantum circuit, configured to process n qubits and d additional qubits, includes: a d-qubit Quantum Fourier Transform circuit configured to apply a d-qubit Quantum Fourier Transform to the d additional qubits; a plurality of parity-fan-out gates controlled by the n qubits and configured to control the additional d qubits, wherein each of the plurality of parity fan-out gates is coupled to a corresponding one a plurality of sets of additional phase gates that also apply phase angles to the d additional qubits, wherein quantum circuit implements a unitary of a bit function and wherein the sets of additional phase gates apply the phase angles to the d additional qubits based on a Walsh-Hadamard Transform of a conversion of the bit function to a binary number; and a d-qubit Inverse Quantum Fourier Transform circuit configured to apply a d-qubit Inverse Quantum Fourier Transform to the d additional qubits.

A mechanism for processing video data is disclosed. The mechanism includes determining to employ side information as input to a filter. The filter can be a bilateral filter (BF) or a Hadamard transform domain filter (HTDF). A conversion is performed between a visual media data and a bitstream based on the filter.

A method of coding is implemented by a decoding device, the method includes: receiving a bitstream, and obtaining from the bitstream, a first luma bit depth value of luma samples of a current picture. The method also includes performing a Hadamard Transform Domain Filtering (HTDF) post reconstruction filtering of a block of the current picture by using a lookup table based on the obtained first luma bit depth value.

The present invention provides a memristor-based image compression system and method, the image compression system comprising: a first storage subarray, a second storage subarray, an input encoding module, an output encoding module, a storage module, and a frequency signal encoding module; image information is converted into a processable signal by the input encoding module, and a Walsh-Hadamard transform operator is mapped to the storage subarray; after completing the image compression forward transform, the frequency information encoding module compresses the obtained frequency information; the image compression system performs an inverse transform of the image compression, and the current value after the image compression inverse transform is output as compressed image information after passing through the output encoding module, thus completing the image compression. The image compression in the present invention uses the Walsh-Hadamard transform. Since the transform operator only includes +1 and -1, only a memristor device with 1-bit precision is required, which greatly reduces the performance requirements of the device and also reduces the complexity of the device control coding.

An orthogonal time sequency multiplexing (OTSM)-based communication method includes generating a plurality of candidate matrices in a delay-time domain by multiplying data in row units in a delay-sequency domain by a plurality of preset scrambled Walsh-Hadamard transform (S-WHT) matrices or generating a plurality of candidate matrices in a delay-time domain by scrambling the data in row units in the delay-sequency domain by using a plurality of preset scramble vectors and then multiplying the scrambled data by the S-WHT matrices, and generating a data matrix in the delay-time domain based on the plurality of candidate matrices.

Infrared spectrometer and method of performing infrared spectrometry. In one embodiment, the method comprises the steps of providing a first single pixel detector sensitive to infrared light in a first spectral range; providing an entrance slit for receiving an infrared light signal; disposing a moveable encoding mask between the entrance slit and the first single pixel detector for encoding based multiplexing, the moveable encoding mask comprising at least three adjacent coding sections along an encoding moving direction thereof, each coding section comprising the same coding pattern in a cyclic manner such that a last encoding step of one encoding section is the same as a first encoding step in a next encoding section step; disposing a dispersion and imaging optics between the entrance slit and the moveable encoding mask for dispersing the infrared signal and for imaging the dispersed infrared signal onto the moveable encoding mask; disposing a collection optics between the moveable encoding mask and the first single pixel detector for collecting an encoding based multiplexed version of the infrared signal onto the first single pixel photodetector; selectively allowing only one of at least first and second bands within the first spectral range to be imaged onto respective ones of the coding sections excluding a first coding section along the encoding moving direction of the moveable encoding mask, in a starting position of the moveable encoding mask; and moving the moveable encoding mask in the encoding moving direction for the encoding based multiplexing.

A quantum circuit, configured to process n qubits and d additional qubits, includes: a d-qubit Quantum Fourier Transform circuit configured to apply a d-qubit Quantum Fourier Transform to the d additional qubits; a plurality of parity-fan-out gates controlled by the n qubits and configured to control the additional d qubits, wherein each of the plurality of parity fan-out gates is coupled to a corresponding one a plurality of sets of additional phase gates that also apply phase angles to the d additional qubits, wherein quantum circuit implements a unitary of a bit function and wherein the sets of additional phase gates apply the phase angles to the d additional qubits based on a Walsh-Hadamard Transform of a conversion of the bit function to a binary number; and a d-qubit Inverse Quantum Fourier Transform circuit configured to apply a d-qubit Inverse Quantum Fourier Transform to the d additional qubits.

This invention provides a method for forcibly partitioning coding units (CUs). The method analyzes the characteristics of the predicted residual data, uses a quadtree, and identifies the consistency of detail richness of high-frequency coefficients in the four sub-blocks after Hadamard transform of the residual data predicted by the coding unit (CU). This determines whether the current CU block size will cause detail loss or detail spread. For CUs with complex residual data (i.e., inconsistent detail richness in the quadtree-partitioned sub-blocks), forced partitioning is performed, and over-partitioning is corrected. This method supplements and corrects the CU decision-making process for pass rate distortion optimization, significantly improving the effects of detail loss and detail spread.

A quantum circuit, configured to process n qubits and d additional qubits, includes: a d-qubit Quantum Fourier Transform circuit configured to apply a d-qubit Quantum Fourier Transform to the d additional qubits; a plurality of parity-fan-out gates controlled by the n qubits and configured to control the additional d qubits, wherein each of the plurality of parity fan-out gates is coupled to a corresponding one a plurality of sets of additional phase gates that also apply phase angles to the d additional qubits, wherein quantum circuit implements a unitary of a bit function and wherein the sets of additional phase gates apply the phase angles to the d additional qubits based on a Walsh-Hadamard Transform of a conversion of the bit function to a binary number, and a d-qubit Inverse Quantum Fourier Transform circuit configured to apply a d-qubit Inverse Quantum Fourier Transform to the d additional qubits.

The invention provides a high-dimensional OTSM implementation method based on generalized index modulation, and relates to the field of wireless communication, and the method comprises the steps: completing the mapping of transmitted digital information through a generalized index modulation technology and a high-dimensional signal constellation diagram, and obtaining an OTSM signal frame of a time-order rate domain through an in-phase orthogonal converter and an OTSM frame generator; a time-order rate domain OTSM signal is converted to a time domain by adopting Walsh-Hadamard transform, and a high-dimensional OTSM transmitter based on generalized index modulation is constructed; the high-dimensional OTSM transmitter is modulated through the generalized index to send digital signals in a wireless channel, and better system performance is achieved with low system implementation complexity. According to the technical scheme, the complexity of system implementation can be effectively reduced, a low system peak-to-average power ratio is achieved, and the energy efficiency of the system and the generalized index modulation reliability are effectively improved through application of the index modulation technology and the high-dimensional signal constellation diagram.

The present invention discloses an active single-pixel imaging method for strongly scattering media based on a conditional generative adversarial network. The method comprises the following steps: designing a single-pixel optical imaging system based on a Hadamard transform to obtain one-dimensional detection signal values ​​of a test target under a series of different scattering conditions and sampling rates; shaping the sampled one-dimensional detection signal values ​​under different scattering conditions and sampling rates into two-dimensional signal images through data averaging preprocessing, which are used as test set images; constructing a deep convolutional conditional generative adversarial network (AMSPI-LSCGAN) based on a least squares loss function model with a compression-excitation block and a residual block; combining the least squares loss with a content loss and an average structural similarity loss as the loss function of the AMSPI-LSCGAN network for training, thereby avoiding training collapse and obtaining reconstructed images with higher fidelity; and inputting the test set images into the trained AMSPI-LSCGAN network to reconstruct the test target.

The invention relates to a color image watermark embedding method and system based on Walsh-Hadamard transform, and the method comprises the steps: carrying out the partitioning of a carrier image, calculating the entropy value of each image block, and carrying out the sorting preprocessing of the image blocks, the method comprises the following steps: selecting a plurality of in-block smooth image blocks as to-be-embedded image blocks, carrying out frequency conversion on the to-be-embedded image blocks through Walsh-Hadamard transform, carrying out secondary partitioning on each converted to-be-embedded image block, selecting an embedding position of a watermark according to a frequency difference value between pixel points after Walsh-Hadamard transform, and carrying out watermark embedding according to the frequency difference value between the pixel points after the Walsh-Hadamard transform. The image attack resistance of the embedded secret-carrying image is improved, and the balance between the unperceivability and robustness of the image watermark based on Walsh-Hadamard transform is improved.

The present disclosure provides a method for encoding input data in an encoder to generate encoded output data, wherein the method comprises instructing to convert the input data to base layer reconstruction data using a down-sampled version of the input data; and generating enhancement layer data providing instructions for enhancing the base layer reconstruction data, wherein the enhancement layer data comprises residual data representing a difference between a reconstruction generated from the base layer reconstruction data and data derived from the input data, wherein the residual data is subjected to at least one transform and at least one data compression method to generate the encoded output data, characterized in that the method comprises arranging the at least one transform to comprise a Hadamard transform to convert the residual data to corresponding transform coefficients, wherein different output elements of the Hadamard transform are arranged as different surfaces to be encoded, and selecting an output order for the transform coefficients in the surfaces to provide a sequence of transform coefficients for the at least one data compression method to generate the encoded output data, wherein the output order is selected to be at least one of a tile raster order and an n x n block "Z" order.

A system according to one aspect of the present disclosure comprises: an input unit for inputting the quantum state of n quantum bits to a quantum computer; and a calculation unit for executing, by controlling the quantum computer, a quantum calculation that involves repeating a unitary transformation in which is included a pseudo-random function taking a bit string of n / 2 bits as input / output, a Hadamard transform, and swapping between high-order n / 2 quantum bits and low-order n / 2 quantum bits on the quantum state a predetermined number of times.

Techniques for training deep neural networks with 4-bit integers are described herein. According to one aspect of the present disclosure, a computer-implemented method for accelerating operations of a deep neural network includes, for each of a plurality of matrix multiplications during forward propagation, obtaining a first matrix by multiplying an activation matrix with a block diagonal Hadamard transform matrix, the activation matrix is an input of the matrix multiplication; obtaining a second matrix by multiplying a weight matrix with the block diagonal Hadamard transform matrix wherein the transpose of the weight matrix is another input of the matrix multiplication; quantizing the first matrix and the second matrix into a 4-bit integer matrix; obtaining a result matrix by multiplying the transpose of the quantized first matrix with the transpose of the quantized second matrix; and inversely quantizing the result matrix into a matrix represented by floating points.

Techniques for training deep neural networks with 4-bit integers are described.According to one aspect of the present disclosure, a computer-implemented method for accelerating operations of a deep neural network comprises: for each of a plurality of matrix multiplications during forward propagation, obtaining a first matrix by multiplying an activation matrix, which is an input to the matrix multiplication, by a block-diagonal Hadamard transform matrix; obtaining a second matrix by multiplying a weighting matrix, one transpose of which is another input to the matrix multiplication, by the block-diagonal Hadamard transform matrix; quantizing the first matrix and the second matrix into 4-bit integer matrices; obtaining a resultant matrix by multiplying the quantized first matrix and the transpose of the quantized second matrix; and dequantizing the resultant matrix into a floating-point matrix.

A video encoding method and apparatus for calculating a sum of absolute transformed difference (SATD) through a folded Hadamard transform circuit includes partitioning a current block into SATD blocks, receiving input data associated with a first block of a first SATD block in a first cycle and input data associated with a second block of the first SATD block in a second cycle, and performing a calculation of the first block by a shared Hadamard transform circuit in the first cycle and a calculation of the second block by the shared Hadamard transform circuit in the second cycle. Each shared Hadamard transform circuit is a first portion of each folded Hadamard transform circuit. The video encoding method and apparatus further perform a calculation of an entire SATD block through a final portion of each folded Hadamard transform circuit to generate a final SATD result for the first SATD block for encoding.

The invention discloses an image-free target invariant moment feature recognition method based on a single-pixel detector, and the method comprises the steps: S1, pre-calculating a Hadamard coefficient of a geometric moment kernel function, calculating and sparsely storing a non-zero Walsh-Hadamard transformation coefficient of the geometric moment kernel function in a Hadamard domain and a corresponding primary function, so as to reduce the storage and calculation overhead; s2, loading a corresponding Hadamard pattern by a modulation module, carrying out image-free target single-pixel measurement, and resolving a target geometric moment in combination with a pre-stored geometric moment kernel function Hadamard coefficient and a single-pixel detection value; and S3, carrying out nonlinear combination on the geometric moment, and constructing an improved Hu invariant moment feature of the target. According to the method, a target image does not need to be reconstructed, a single-pixel detection technology is fused, Hadamard domain sparsity optimization calculation is fully utilized, and high-precision feature extraction at a low sampling rate is realized; the proposed improved Hu invariant moment significantly improves robustness to rotation, translation, scaling, modulator size, and detection gain variations.

This invention belongs to the field of wireless communication and array signal processing technology, specifically relating to a low-precision phase shifter codebook design method suitable for near-field MIMO. Addressing the problems of large size, high computational complexity, and strong dependence on high-precision phase shifters in traditional angle-range domain codebooks in near-field hybrid precoding, this invention constructs a low-precision analog domain codebook based on the complex Hadamard transform of conjugate symmetric sequences, ensuring that codeword elements are taken only from [specific data type], thus supporting 2-bit phase shifter implementation. Furthermore, it improves beam coverage density by constructing an overcomplete codebook and combines it with a virtual subarray structure to perform beam selection at the subarray scale to adapt to the propagation characteristics of near-field spherical waves. This invention achieves good precoding performance under near-field and spatially non-stationary channel conditions while reducing hardware cost and power consumption, making it suitable for ultra-large-scale MIMO and near-field hybrid precoding systems.

A computer-implemented method is presented for performing a computation with a neural network. The method includes: receiving a first input patch of data; applying a Walsh-Hadamard transform to the input patch to yield a transformed input patch in a transformed domain; computing an element-wise product of the transformed input patch and a kernel of the neural network; applying an inverse Walsh-Hadamard transform to the element-wise product to yield an intermediate matrix; and creating a first output patch from the intermediate matrix, where the size of the first output patch is smaller than the intermediate matrix.

This application discloses a deployment method, electronic device, and readable storage medium for a pre-trained language model, relating to the field of pre-trained language model technology. The method includes: determining the initial weight matrix and Hessian matrix of the initial pre-trained language model to be quantized; the initial weight matrix is ​​used to characterize the model information topology of the initial pre-trained language model, and the Hessian matrix is ​​used to characterize the sensitive weights in the initial weight matrix; generating a sparse transformation matrix based on a subsampled random Hadamard transform algorithm; performing incoherent processing on the initial weight matrix using the sparse transformation matrix to obtain a processed optimized weight matrix; performing adaptive rounding based on the optimized weight matrix and the Hessian matrix to obtain a quantized weight matrix; determining the quantized target language model based on the quantized weight matrix, and sending it to the deployment device. This application solves the problem of low efficiency in deploying pre-trained language models, achieving the technical effect of improving the efficiency of deploying pre-trained language models.