271 results about "Quantization matrix" patented technology

The invention discloses a quantification method of an artificial intelligence model, a processing system of the artificial intelligence model, a quantification unit and computing equipment. A quantization unit may divide a quantization matrix into a plurality of block quantization matrices applicable to processing granularities of respective computing units of a computing device running an AI model based on the processing granularities of the respective computing units. The quantization unit can perform quantization processing on the block quantization matrix to obtain the quantized block quantization matrix, and the quantized block quantization matrix is stored in the memory, so that the memory occupied by the AI model can be reduced. The quantization matrix can be a weight matrix or an activation matrix. In the quantization process of the quantization unit, quantization is carried out based on the processing granularity of each computing unit of the computing device running the AI model, outlier statistics does not need to be carried out based on the tensor level, and real-time quantization can be carried out on the weight matrix and the activation matrix by the AI model.

The invention provides a model training method and device, equipment and a storage medium, and relates to the technical field of computers, in particular to the technical field of neural network models and model training. The specific implementation scheme is as follows: a calculation unit executes quantization matrix multiplication based on Hadamard pre-transformation on an activation tensor and a weight tensor of a target model stored in a memory so as to generate an output tensor of a linear layer based on a low-precision tensor with smaller data bit width; using the output tensor and a subsequent network layer of the target model to complete forward propagation so as to obtain a loss value; and according to the loss value, updating model parameters of the target model stored in a memory through a back propagation algorithm. By means of the technical scheme, on the premise that the model training precision is guaranteed, memory resource occupation and the calculation amount in the calculation process can be remarkably reduced, and the training cost is reduced.

The invention discloses a network security situation awareness method based on artificial intelligence, and the method comprises the following steps: collecting multi-source heterogeneous data, and generating a standardized data set; spatial-temporal feature decoupling is carried out, and spatial-temporal dimension features are separated; fusing the time-space cross attention, and outputting a fused time-space feature vector; constructing a causal inference engine, and outputting a dynamic causal graph and an anti-fact inference result set; constructing a dynamic risk propagation model, and outputting a whole asset risk value matrix and a risk propagation path diagram; generating a situation quantization matrix, constructing an adversarial training decision network, and outputting a defense strategy set verified by adversarial training; automatically generating a strategy; and a man-machine cooperative verification closed loop is realized. According to the method, dynamic reconstruction of a threat propagation path is realized through spatial-temporal feature decoupling and a causal reasoning engine, and a risk positioning error is reduced; and the adversarial training decision network is combined, so that the misjudgment rate of the defense strategy in the simulation APT attack test is reduced.

The embodiment of the invention provides a data processing method and device for large model parameters, equipment and a medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: obtaining a quantization matrix and a meta-parameter of a target large model from different partitions of a first data storage area of a memory, reading memory access granularity representing the data processing capability of a processor, and determining the size of a filling area between the insertion meta-parameter and the quantization matrix based on the memory access granularity; dividing a second data storage area different from the first data storage area in the memory, and sequentially writing the meta-parameter, preset filling information matched with the filling area in size and the quantization matrix into the second data storage area to obtain an initial data atomic block; and continuously arranging all the initial data atomic blocks in the second data storage area to obtain a target data atomic block. By optimizing the data storage mode, the memory bandwidth utilization rate of the processor in the reasoning process is improved, and then the data calculation efficiency of the processor is improved.

The invention discloses an automobile wire harness detection device based on artificial intelligence and machine vision, and relates to the technical field of visual inspection, and the device comprises a film covering module which enables an automobile wire harness to be immersed in a control groove of magnetic fluid to form a nano film; the marking module projects interference fringes to the nano-film through a laser interference projector, and marks timestamps and capture numbers of fringe images; the image module is used for applying magnetic field pulses in the timestamps of the adjacent stripe images to generate a turbulent flow image sequence; the quantization module is used for respectively carrying out interference fringe phase analysis on the turbulence images in the turbulence image sequence to generate a curvature quantization matrix; the defect module is used for analyzing the curvature quantization matrix through a pre-trained convolutional neural network and identifying a defect type and a coordinate position; and the result module is used for superposing the defect type and the coordinate position to the stripe image to generate a defect distribution thermodynamic diagram. According to the invention, high-precision intelligent defect identification of wire harness detection is realized through cooperation of the nano-film and the convolutional neural network.

The invention relates to the technical field of optical equipment testing, in particular to a radiation therapy laser positioning system performance testing method, which comprises the following steps of: testing first positioning data of optical positioning equipment in a load state, constructing a body weight-optical response correlation model, and analyzing the body weight-optical response correlation model to obtain static optical equipment testing data; a tester slightly moves along a preset track, a response signal of the laser photosensitive sensor and space coordinates of the laser tracker are collected to generate second optical positioning data, and a motion-optical response coupling model is constructed to analyze the second optical positioning data to obtain dynamic optical equipment test data; a multi-factor deviation prediction model is constructed to analyze optical interference signals, phase detection data and environment temperature and humidity parameters to generate optical equipment compensation parameters, and static optical equipment test data, dynamic optical equipment test data and the optical equipment compensation parameters are analyzed; and calculating and generating an optical calibration scene fitness score, a positioning precision quantization matrix and an optical equipment acceptance reference value comprehensive evaluation report.

The invention provides a model quantification method and device, a model reasoning method and device, equipment, a medium and a product. Performing sensitivity evaluation on each network channel based on the weight matrix to obtain the channel sensitivity of each network channel; based on the channel sensitivity of the network channel corresponding to each element in the weight matrix, rearranging and dividing each element in the weight matrix to obtain a plurality of weight sub-matrixes; based on the channel sensitivity of the network channel corresponding to each element in each weight sub-matrix, allocating quantitative resources to each weight sub-matrix; performing quantization processing on the weight sub-matrixes based on the quantization resources to obtain a plurality of quantization sub-matrixes; and correcting the plurality of quantization sub-matrixes to obtain a target quantization matrix. Elements in a weight matrix can be accurately and orderly gathered according to channel sensitivity, different weight parameters are finely classified, different weight sub-matrixes are compatible for quantification, the problem of weight heterogeneity caused by multiple modes is solved, and the accuracy of model quantification is improved.

The invention relates to the field of image encryption, in particular to a chaos image encryption method based on DNA coding, which comprises the following steps: S1, calculating a hash value of a plaintext grayscale image by using a hash algorithm, and generating an initial key based on the hash value, S2, iterating a chaos sequence by using the initial key as an initial value and a control parameter of an iteration variable of two-dimensional chaos mapping 2D-CSMM, performing segmentation and quantization processing on the chaotic sequence to obtain a quantization matrix; performing reduction XOR operation on the quantization matrix to obtain a bit matrix; performing scrambling processing on a plaintext grayscale image based on the quantization matrix and the bit matrix to obtain a ciphertext image; s5, performing DNA diffusion processing on the ciphertext image based on the quantization matrix and the bit matrix to obtain an encrypted image; s6, decrypting the encrypted image according to the inverse process of image encryption to obtain a decrypted image; according to the encryption method provided by the invention, the security of image encryption is greatly improved, and the security performance and the encryption effect are good.

An encoding apparatus comprises a quantizer quantizing transform coefficients in a block in an image using a first quantization matrix, and an encoder encoding the first quantization matrix, wherein, for encoding the first quantization matrix, if using a first mode, the encoder encodes information indicating that the first quantization matrix is the same as a second quantization matrix having already been encoded, if using a second mode, encodes a difference between an element of the first quantization matrix and an element of the second quantization matrix, and if using a third mode, encodes a difference between an element and another element within the first quantization matrix, wherein a code amount of information representing the first mode is used is less than that of information representing that the second mode is used.

The invention discloses a federated learning method and device using high-order sub-models to assist training, and the method comprises the steps: enhancing the reasoning accuracy of a low-order sub-model through integrating the decision information of multiple high-order sub-models; the low-order sub-model can assist in training by means of decision information of the high-order sub-model and a quantization matrix of the highest-order sub-model based on the feature distance, and the training effect is improved; a low-order sub-model is optimized by reducing target loss (information entropy) of target category confidence, and the problem of information offset caused by re-fitting of target category confidence of a high-order sub-model in traditional self-distillation is avoided; unifying feature dimensions of all levels of sub-models by utilizing comparative learning, constructing a quantization matrix of the highest-order sub-model based on feature distance as supervision information, and guiding the low-order sub-model to approach the quantization matrix of the highest-order sub-model by minimizing the distribution difference between the low-order sub-model and the distance matrix of the highest-order sub-model, so as to realize the supervision of the highest-order sub-model. And a low-order sub-model is guided to mine semantic relationships in different images, so that the generalization ability of the model is improved.

An encoding apparatus comprises a quantizer quantizing transform coefficients in a block in an image using a first quantization matrix, and an encoder encoding the first quantization matrix, wherein, for encoding the first quantization matrix, if using a first mode, the encoder encodes information indicating that the first quantization matrix is the same as a second quantization matrix having already been encoded, if using a second mode, encodes a difference between an element of the first quantization matrix and an element of the second quantization matrix, and if using a third mode, encodes a difference between an element and another element within the first quantization matrix, wherein a code amount of information representing the first mode is used is less than that of information representing that the second mode is used.

An encoding apparatus comprises a quantizer quantizing transform coefficients in a block in an image using a first quantization matrix, and an encoder encoding the first quantization matrix, wherein, for encoding the first quantization matrix, if using a first mode, the encoder encodes information indicating that the first quantization matrix is the same as a second quantization matrix having already been encoded, if using a second mode, encodes a difference between an element of the first quantization matrix and an element of the second quantization matrix, and if using a third mode, encodes a difference between an element and another element within the first quantization matrix, wherein a code amount of information representing the first mode is used is less than that of information representing that the second mode is used.

The invention discloses an image compression method and system based on edge gradient self-adaption, and the method comprises the steps: converting an input image into a YCbCr color space, calculating an edge gradient value of a brightness component, and employing iteration processing including a feedback mechanism: dividing an edge region and a non-edge region based on a current self-adaption edge gradient threshold value, the blocks are divided according to different sizes; dynamically updating an edge gradient threshold according to a block statistics result, and adaptively optimizing region division through multiple iterations; and finally, based on the optimized blocking result, processing the edge region and the non-edge region by adopting different quantization strategies, generating a quantization matrix, and then carrying out entropy coding to output a compressed code stream. According to the method, adaptive adjustment of compression parameters is realized through an iterative optimization mechanism, image edge and texture details are effectively reserved while a high compression rate is ensured, and the subjective visual quality of a compressed image is remarkably improved.

An encoding apparatus comprises a quantizer quantizing transform coefficients in a block in an image using a first quantization matrix, and an encoder encoding the first quantization matrix, wherein, for encoding the first quantization matrix, if using a first mode, the encoder encodes information indicating that the first quantization matrix is the same as a second quantization matrix having already been encoded, if using a second mode, encodes a difference between an element of the first quantization matrix and an element of the second quantization matrix, and if using a third mode, encodes a difference between an element and another element within the first quantization matrix, wherein a code amount of information representing the first mode is used is less than that of information representing that the second mode is used.

The invention relates to the technical field of video image compression, in particular to a video image compression algorithm based on an H.266 / VVC standard, which comprises the steps of generating a visual attention thermodynamic diagram through time-space domain saliency detection, adjusting a coding tree unit QP value in a partitioned manner to realize adaptive quantization, and designing an asymmetric quantization matrix to optimize a transformation quantization effect. And the generative adversarial network is used to enhance the quality of the reconstructed frame. According to the method, the blocking effect of a complex texture region can be effectively inhibited, the quantization distortion of a human eye sensitive region is reduced, the edge preserving capability is improved, the subjective visual quality and compression efficiency of the video are remarkably improved, and the high-definition video transmission requirement is met.

Techniques related to adaptive quantization matrix selection using machine learning for video coding are discussed. Such techniques include applying a machine learning model to generate an estimated quantization parameter for a frame and selecting a set of quantization matrices for encode of the frame from a number of sets of quantization matrices based on the estimated quantization parameter.

Different quantization matrices may be transmitted corresponding to different block sizes, color components and prediction modes. To more efficiently signal the coefficients of the quantization matrices, in one implementation, a unified matrix identifier matrixId is used, based on a size identifier (sizeId) that relates to CU size with larger sizes listed first, and a matrix type (matrixTypeId) with luma QMs listed first. For example, the unified identifier is derived as: matrixId=N*sizeId+matrixTypeId, where N is the number of possible type identifiers, e.g., N=6. This single identifier allows referring to any previously transmitted matrix when using prediction (copy), and transmitting larger matrices first avoids interpolation in the prediction process. When a block uses the Intra Block Copy prediction mode, QM identifier may be derived as if the block uses the INTER prediction mode.

The invention discloses a wireless sensing and communication method based on a super-diagonal reconfigurable intelligent surface, and relates to the technical field of communication sensing, and the method comprises the steps: building an ISAC system model, and carrying out the tracking and positioning of a sensing target through an echo signal received by a sensing communication base station based on the ISAC system model, proposing an optimization problem of the minimization joint posterior Cramer-Rao bound according to a tracking and positioning result; based on the optimization problem, a first sub-problem used for optimizing beam forming of the sensing communication base station and a second sub-problem used for optimizing a super-diagonal reconfigurable intelligent surface reflection coefficient are decomposed, and the first sub-problem is converted into a convex problem to be solved by using a positive semidefinite relaxation algorithm; and solving the second sub-problem by using an augmented Lagrange multiplier method and a penalty dual decomposition framework vectorization matrix. According to the invention, through joint optimization of the beam forming of the sensing communication base station and the reflection coefficient of the super-diagonal reconfigurable intelligent surface, improvement of sensing target parameter estimation precision is realized.

Encoding an extended-reality (XR) video frame may include obtaining an XR video frame comprising a background image and a virtual object; obtaining, from an image renderer, a first region of the background image over which the virtual object is overlaid; dividing the XR video frame into a virtual region and a real region, wherein the virtual region comprises the first region of the background image and the virtual object and the real region comprises a second region of the background image; determining, for the virtual region, a corresponding first quantization parameter based on an initial quantization parameter associated with virtual regions; determining, for the real region, a corresponding second quantization parameter based on an initial quantization parameter associated with real regions; and encoding the virtual region based on the corresponding first quantization parameter and the real region based on the corresponding second quantization parameter.

A video coding system for a machine employs adaptive quantization of a frequency response based on a machine model of received data. The encoder receives a machine model of the machine-based system and generates a frequency importance map from the machine model. An adjustment matrix is then generated based on the frequency importance map and used to adjust the coefficients of the default quantization matrix. Video data is quantized using the adjusted quantization matrix and encoded in a bitstream for transmission to a decoder site. At the decoder site, the decoder may extract parameters of the adjusted quantization matrix from the bitstream, or calculate the adjusted quantization matrix using a machine model to inverse quantize the received data for machine consumption.

The invention belongs to the technical field of vehicle-road cloud integrated systems, and discloses a method for analyzing functional requirements of a vehicle-road cloud integrated system based on LLM, and the method comprises the following steps: S1, collecting and preprocessing basic data; s2, structured extraction of VRCIS requirements: semantic analysis is performed on the design requirement document based on LLM, a data structure is converted, and functional requirements and non-functional requirements are extracted; s3, constructing a VRCIS demand and benefit-critical person matrix: based on the output of the S2, mapping the function demand of each VRCIS benefit-critical person and the system design key dimension, and generating a standardized quantization matrix which can be identified by a machine; s4, quantitative analysis of demand priorities: identifying the demand priorities through quantitative analysis by utilizing the benefit-critical person matrix constructed in the S3 and fusing cost-benefit analysis; and S5, outputting a demand analysis result. The method is high in analysis efficiency, and the problems that in the analysis process, demand decomposition is not in place, the demand priority is not clear, and a general demand template cannot fully express VRCIS complexity are solved.

An encoder, includes: memory; and circuitry coupled to the memory, in which in a first encoding process, the circuitry quantizes a current block using a first quantization matrix in both of a first case where the circuitry performs primary transform and the circuitry does not perform secondary transform and a second case where the circuitry performs both the primary transform and the secondary transform, in a second encoding process, the circuitry quantizes the current block using a second quantization matrix in a third case where the circuitry performs primary transform and the circuitry does not perform secondary transform, and in a third encoding process, the circuitry quantizes the current block without using a quantization matrix in both of a fourth case where both the primary transform and the secondary transform are skipped and a fifth case where both the primary transform and the secondary transform are performed.

A system and method for energy-efficient approximate digital JPEG and MJPEG-compression. The system includes a controller unit to control a processing loop for processing image blocks based on a comparison of a current image block to a previous image block. The system includes a quantization unit configured to quantize the frequency domain representation using an approximate quantization process and a quantization (Q) matrix. The quantization unit is configured to: identify, a nearest power of two value for each element of the quantization matrix; generate an updated Q matrix by assigning each element of the quantization matrix with the identified nearest power of two value; and shift each element of the updated Q matrix by a number of bits to generate a quantized frequency domain representation. The number of bits corresponds to the identified nearest power of two for the corresponding element of the updated Q matrix.

Disclosed is an operation method for executing an addition operation on quantized matrices. This operation method comprises the steps in which: an NPU matrix multiplication unit executes a matrix multiplication operation on a first matrix and a second matrix, wherein the first matrix has a size of R0*2C0 and is obtained by concatenating a first input tensor and a second input tensor, each of which is composed of quantized values and has a size of R0*C0, and the second matrix is obtained by concatenating a first diagonal matrix having a size of C0*C0 and having all main diagonal element values equal to integer N1 and a second diagonal matrix having a size of C0*C0 and having all main diagonal element values equal to integer N2; and an NPU vector processing unit calculates elements of a non-quantized tensor obtainable by adding the second input tensor to the first input tensor, by adding real number B to a product of real number R and a third matrix calculated by the matrix multiplication operation.

An image encoding apparatus for encoding an image comprises a generation unit configured to generate, from a first quantization matrix having a size of N×N (N is a positive integer), a second quantization matrix having a size of P×Q (P and Q are positive integers which satisfy P<N<Q, and the size of P×Q indicates a size of P in a horizontal direction and a size of Q in a vertical direction), and a quantization unit configured to quantize transformation coefficients in a sub-block having a size corresponding to the size of P×Q using the second quantization matrix.

The invention relates to the technical field of artificial intelligence model optimization, in particular to a large model reasoning acceleration method and device for packet sensing quantization and residual correction, and the method comprises the steps: carrying out the statistical analysis of the weight and activation of each layer of a large model, and generating a channel feature matrix; constructing a learnable grouping mapping matrix, dividing channels into different groups, and dynamically distributing quantized bit width to obtain a grouping weight matrix; on the basis of the channel weight-activation joint sensitivity, calculating each grouping error contribution on line by using a small prediction model, and adjusting a grouping weight matrix to generate an optimized weight matrix; constructing an error control matrix to dynamically adjust the quantization error along the propagation path, and generating a correction matrix; dynamically adjusting the sparse rate and the quantization precision according to the channel feature matrix and the hardware constraint by combining a structured sparse strategy, and generating a sparse quantization matrix; in the reasoning process, model reasoning is carried out according to the correction matrix and the sparse quantization matrix, and the weight matrix and the correction matrix are updated and optimized in a closed-loop mode.

The application discloses a network security situation awareness method based on artificial intelligence, comprising the following steps: collecting multi-source heterogeneous data to generate a standardized data set; decoupling space-time characteristics to separate space-time dimension characteristics; fusing space-time cross attention to output fused space-time characteristic vectors; constructing a causal reasoning engine to output a dynamic causal diagram and a counterfactual reasoning result set; constructing a dynamic risk propagation model to output a full asset risk value matrix and a risk propagation path diagram; generating a situation quantization matrix, constructing an adversarial training decision network, and outputting a defense strategy set verified by adversarial training; automatically generating a strategy; and man-machine collaborative verification closed loop. Through space-time characteristic decoupling and the causal reasoning engine, the application realizes dynamic reconstruction of a threat propagation path, reduces risk positioning error; in combination with the adversarial training decision network, the application makes the defense strategy have a reduced misjudgment rate in the simulation APT attack test.

The invention discloses an image compression method and device, and the method comprises the steps: comparing a threshold matrix with a frequency domain coefficient matrix, so as to calculate the number of effective coefficients of each frequency point in each frequency domain coefficient matrix, forming an effective coefficient matrix, and achieving the independent calculation of the scaling of each frequency point based on the effective coefficient matrix and a scaling calculation formula. According to the method, a frequency domain coefficient matrix is obtained, a corresponding scaling matrix is formed, a corresponding optimized quantization matrix is generated according to the scaling matrix, the frequency domain coefficient matrix is quantized by using the optimized quantization matrix, scanning, data rearrangement and entropy coding operation are executed, and image compression is completed.

Disclosed are an image encoding / decoding method and a non-transitory computer-readable recording medium. The image decoding method includes: determining whether transform skipping can be applied to a current block; determining whether the current block is a transform skipping block based on the above determination; deriving a scaling factor of the current block depending on whether the current block is a transform skipping block; performing inverse quantization by scaling quantized transform coefficients of the current block based on the scaling factor; obtaining residual samples of the current block by selectively performing inverse transform on the scaled transform coefficients; performing prediction on the current block to generate prediction samples of the current block; and reconstructing the current block based on the residual samples and the prediction samples. When the current block is not a transform skipping block, the scaling factor is derived based on a quantization matrix and locations of transform coefficients within the current block. When the current block is a transform skipping block, the scaling factor is derived to be equal to a fixed constant value. The transform skipping block is specified based on information indicating whether an inverse transform is applied to the current block.

Disclosed herein is a video decoding method including determining one or more adaptation parameter sets including a quantization matrix set including a plurality of quantization matrices, determining an adaptation parameter set including a quantization matrix set applied to a current picture or a current slice from among the one or more adaptation parameter sets, dequantizing transform coefficients of a current block of a current picture or a current slice based on the quantization matrix set of the determined adaptation parameter set, and reconstructing the current block based on the dequantized transform coefficients, wherein the adaptation parameter set includes coding information applied to one or more pictures or slices.