453 results about "Inverse quantization" patented technology

The present disclosure provides a variable-bit-rate image compression method and system, an apparatus, a terminal, and a storage medium. The variable-bit-rate image compression method includes: obtaining an initial feature map from a to-be-encoded image; quantizing the initial feature map by a dead-zone quantizer; performing entropy encoding on the quantized feature map and hyper-prior information to obtain a compressed bit-stream; performing entropy decoding on the compressed bit-stream, and recovering quantized hyper-prior information and the quantized feature map; performing inverse quantization on the quantized feature map to obtain a reconstructed feature map; obtaining a reconstructed image from the reconstructed feature map; and adjusting quantization and inverse quantization parameters according to a target bit-rate or target distortion. The present disclosure provides a precise bit-rate control solution, makes the bit-rate of the compressed bit-stream better adapt to the dynamic change of a network bandwidth, and has an extremely high actual application value.

The invention relates to the technical field of voice processing, can be applied to service system platforms of medical health, financial science and technology, communication and the like, and discloses an audio compression and reconstruction method, device, equipment and medium. Residual vector quantization and inverse quantization are carried out to generate basic spectrum features; and compensating a high-frequency component by adopting a spectrum expansion network, predicting a time domain error in combination with a long-short-term memory network, generating a time domain residual compensation signal, and superposing the time domain residual compensation signal with the reconstructed time domain audio signal to obtain a target audio signal. Through residual vector quantization and neural network modeling, the audio reconstruction quality under low-bit-rate compression is improved, spectrum detail reservation and time domain error compensation are optimized, meanwhile, the calculation complexity is reduced, and the method is suitable for high-sampling-rate and resource-limited scenes.

Provided in the present invention is a parameter-efficient large-language fine-tuning federated learning framework, comprising the following steps: performing modeling on LoRA adapters of different edge clouds; since different weights exhibit different average performances on the LoRA adapters, using singular values to quantify the importance of the weights, and therefore, before each round of independent training of the LoRA adapters using N edge clouds, using a matrix singular value to decompose a BA matrix in the LoRA adapter for each trainable weight; configuring heterogeneous LoRA adapters on the basis of the importance of the weights; and using different numbers of quantization bits to quantize a pre-trained model, and performing high-precision inverse quantization on the pre-trained model only when matrix multiplication is executed, wherein the pre-trained model is quantized to the maximum number of quantization bits on the basis of the memory budget of the edge clouds. The present invention has the following beneficial effects: the present invention determines the optimal fine-tuning model structure, thereby improving the performance of LLM fine-tuning, and adapts to heterogeneous and resource-constrained edge clouds.

The invention discloses mixed precision multiply-add operation hardware, which comprises an input processing unit used for obtaining standard symbols, standard indexes and standard mantissas according to a pre-configured input format and unpacked input multiply-add operation factors, and obtaining input and output special value detection results; the floating point multiplier-adder is used for calculating an operation result after the operand multiplies the scaling factor and the offset is accumulated according to the symbol, the exponent and the mantissa of each multiplication-addition operation factor; and the output processing unit is used for obtaining a target format result according to a pre-configured output format, a rounding type and a normalization and rounding operation result, packaging the target format result according to a detection result and outputting the target format result, and by multiplexing a data path of the floating-point multiplier-adder, the chip area is saved, format conversion is integrated, a data stream is simplified, and the data processing efficiency is improved. The requirements for automatic reconstruction of the multiply-add operation unit according to input and output formats and quantization and inverse quantization of floating points and integers are met, and the reuse rate of computing resources is increased.

A device for decoding encoded dynamic mesh data determines a set of quantized integer coefficient values for displacement vectors of the encoded dynamic mesh data; determines a quantization parameter based on one or more syntax elements included in a displacement sub-bitstream of the encoded dynamic mesh data; inverse quantizes, based on the quantization parameter, the set of quantized integer coefficient values to determine a set of fixed-point dequantized coefficient values; determines a set of fixed-point transformed coefficient values based on the set of fixed-point dequantized coefficient values; converts the set of fixed-point transformed coefficient values to a set of floating-point transformed coefficient values; inverse transforms the set of floating-point transformed coefficient values to determine a set of reconstructed displacement vectors; and determines a reconstructed deformed mesh based on the set of reconstructed displacement vectors.

The invention relates to the technical field of voice processing, can be applied to business scenes of financial science and technology, medical health and the like, and discloses a voice signal compression method, device, equipment and medium, which comprises the following steps: executing Fourier transform on an initial voice signal to extract an amplitude spectrum and a phase spectrum, respectively processing the amplitude spectrum and the phase spectrum to generate features, and splicing the features to obtain a compressed voice signal; the method comprises the following steps: carrying out quantization by adopting a residual vectorization mode to generate a compressed feature vector, executing entropy coding to obtain a compressed code stream, recovering the compressed feature vector at a decoding side, reconstructing a spliced feature through residual reverse quantization, enhancing feature expression through up-sampling operation, and restoring a voice signal through inverse Fourier transform. According to the method, a dual-path processing structure of amplitude features and phase features is constructed, feature compression is realized in combination with residual vectorization and entropy coding, inverse quantization and up-sampling enhancement are executed in a decoding stage, a spectrum energy structure and phase continuity are effectively reserved, and reconstruction precision and fidelity of voice signals are improved under the condition of low bit rate.

The invention provides a sparse representation-based low-bit-rate video coding method, which comprises the following steps of: after a coding end acquires an input video, carrying out frame segmentation sampling on the input video to obtain a video frame sequence, and carrying out block segmentation on elements in the video frame sequence to obtain an image block sequence; processing elements in the image block sequence based on a dictionary learning algorithm and a sparse representation algorithm of gradient sparsity to obtain a sparse representation coefficient; sequentially performing quantization processing and entropy coding processing on the sparse representation system to obtain a compressed bit stream, and storing and transmitting the compressed bit stream; obtaining a compressed bit stream, and sequentially performing entropy decoding processing and inverse quantization processing on the bit stream to obtain an original sparse representation coefficient; processing the original sparse representation coefficient based on a dictionary learning algorithm and a sparse decoding algorithm of gradient sparsity to obtain a reconstructed image block; merging the plurality of reconstructed image blocks to obtain a reconstructed video frame; and carrying out frame combination on the plurality of reconstructed video frames to obtain a reconstructed input video.

A method and an apparatus are disclosed for patch-based mesh coding. In the disclosed embodiments, a mesh decoding device decodes a bitstream to reconstruct patch information and a patch-based base mesh. The mesh decoding device reconstructs base mesh vertices and connectivity by using the patch information and the patch-based base mesh. The mesh decoding device generates predicted vertices and connectivity based on the reconstructed base mesh vertices and connectivity. The mesh decoding device decodes a bitstream to reconstruct a transform-coefficient image, and reconstructs vector differences of vertices by unpacking, inverse quantizing and inverse transforming the transform-coefficient image. The mesh decoding device adds the predicted vertices and the vector differences to reconstruct mesh vertices and connectivity.

The invention relates to the technical field of integrated circuit design and artificial intelligence hardware acceleration, and particularly discloses a quantization precision adaptive switching hardware acceleration architecture. Hardware is divided into a static area and a dynamic area, wherein the static area comprises a pre-processing module, a post-processing module, a control module and a reconfigurable logic loading control block which run all the time; the dynamic region bears a quantization and inverse quantization module composed of a plurality of reconfigurable logic units. The method is characterized in that a model precision instruction is analyzed through a control module, and a reconfigurable logic file with corresponding precision is dynamically loaded to a dynamic area from external starting equipment by a loading control block according to an internal address mapping table, so that seamless switching of different quantization precision is realized on the hardware level, and the whole chip does not need to be burnt again. According to the design, through a modularized and reconfigurable hardware structure, the limitation of fixed-precision hardware in coping with diversified AI models is effectively solved, and the deployment flexibility and energy efficiency are remarkably improved.

The invention discloses a code rate adaptive image compression method and system, and the method comprises the steps: obtaining an input image, and calculating the local information entropy of the input image to generate a spatial adaptive mask; obtaining target code rate and actual code rate information to generate a code rate control signal; and dynamically selecting a calculation path by using a mask through a dynamic reversible neural network, and executing reversible quantization transformation according to the control signal to obtain a compressed bit stream and actual code rate information for closed-loop feedback. According to the method, a symmetric reversible framework with shared parameters is constructed, so that the model storage requirement is reduced; the calculation depth of the network is dynamically adjusted through entropy sensing gating, and the overall calculation amount is reduced; and a composite control and PID optimization mechanism is adopted, so that continuous and smooth code rate control is realized. The deployment and application efficiency of the compression system in a resource limited scene can be effectively improved.

The invention provides a hyperspectral remote sensing image compression method based on attention and quantization coding optimization, and the method comprises the steps: carrying out a network model training process: carrying out the processing, cutting and enhancement of hyperspectral remote sensing image data, and constructing a sample set for training; extracting low-dimensional feature representation of the sample data by using a lightweight encoder, wherein the encoder integrates a convolutional layer and a spectrum multi-head self-attention module; a decoder fusing a space-spectrum attention mechanism is adopted to gradually reconstruct a hyperspectral remote sensing image from low-dimensional features; optimizing coding and decoding model parameters through a combined loss function; a quantization coding two-stage compression process: performing adaptive quantization on the features, and mapping the floating point type features into discrete integers based on a logarithm mapping strategy; performing two-stage coding compression on the quantized features; and recovering feature representation through decoding and inverse quantization, and inputting a trained decoder network to reconstruct a hyperspectral remote sensing image.

The disclosure provides a method for picture processing and a storage medium. The method includes the following. A bitstream is decoded to obtain a quantization coefficient of a current picture block. A quantization parameter corresponding to the current picture block is determined, and a transform coefficient of the current picture block is obtained by performing inverse quantization on the quantization coefficient based on the quantization parameter. A reconstructed picture block of the current picture block is determined according to the transform coefficient. An enhanced picture block is obtained by performing quality enhancement on the reconstructed picture block based on the quantization parameter.

In some embodiments, a method receives a quantized latent representation of an image in a latent space. The image is encoded into a representation in the latent space and quantized to generate the quantized latent representation. A time step parameter is received that is generated based on the representation. The method performs an inverse quantization process to generate a reconstructed representation. A diffusion model performs a denoising process for a number of iterations based on the time step parameter to remove noise from the reconstructed representation to generate a denoised reconstructed representation. The denoised reconstructed representation is decoded into a reconstructed image.

The invention relates to the technical field of steel coil detection, and discloses a multi-mode-based steel coil surface defect detection method, device and equipment and a medium, and the method comprises the following steps: obtaining an input image containing a steel coil surface and a task description corresponding to the input image; performing code conversion on the input image and the task description, and determining a discrete sequence; inputting the discrete sequence into an encoder and decoder model to obtain a task output sequence; and performing inverse quantization reasoning according to the task output sequence, and determining a detection frame, a mask and defect description information of an input image so as to complete surface defect detection of the steel coil. According to the method, samples of different defect types can be rapidly obtained through the sample pair formed by the input image and the task description, rapid iteration of model training is facilitated, and meanwhile, through the multi-modal image set, the detection frame, the mask and the defect description information are directly output, the steel coil surface defect detection precision is improved, and the detection application range is widened.

Techniques are described for decoding video data. A video decoder may determine chroma blocks in a chroma quantization group (QG) of the video data, determine a quantization parameter predictor that is the same for each of the chroma blocks of the chroma QG, determine an offset value that is the same for two or more of the chroma blocks of the chroma QG, determine a quantization parameter value for each of the two or more of the chroma blocks in the chroma QG based on the quantization parameter predictor and the offset value inverse quantize coefficients of one or more residual blocks for the chroma blocks based on the determined quantization parameter value, generate the one or more residual blocks based on the inverse quantized coefficients, and reconstruct the chroma blocks based on the one or more residual blocks.

A quantum computing device is formed using a first chip (302) and a second chip (306). The first chip has a first substrate (303), a first set of pads (312A, B), and a set of Josephson junctions (304) disposed on the first substrate. The second chip has a second substrate (307), a second set of pads (308) disposed on the second substrate opposite the first set of pads, and a second layer (310) formed on a subset of the second set of pads. The second layer is configured to bond the first chip and the second chip. The subset of the second set of pads corresponds to a subset of the set of Josephson junctions selected to avoid frequency conflicts between qubits in a set of qubits. Qubits are formed using Josephson junctions in the subset of Josephson junctions and another Josephson junction not in the subset that is caused to be unavailable for forming qubits.

The invention provides a model reasoning method, device and equipment and a storage medium, and relates to the technical field of computers, in particular to the artificial intelligence fields of neural networks, model reasoning, model quantification and the like. According to the specific implementation scheme, a to-be-processed quantization weight matrix of a model is read into a register of a processor from a cache of the processor; performing inverse quantization fusion calculation based on elements of the quantization weight matrix to be processed in the register and an activation matrix of the model to obtain a reasoning result of a current operator of the model; and writing the reasoning result of the current operator into a memory of the processor for operators after the current operator to use. The calculation amount in the reasoning process can be reduced, and the calculation speed and efficiency of reasoning can be improved.

A data processing method, applied to the field of artificial intelligence, and comprising: acquiring first data, wherein the first data is obtained by means of a machine learning model on the basis of first input data; performing non-uniform quantization processing on the first data to obtain first compressed data, and storing the first compressed data in a memory; and reading the first compressed data from the memory, and performing inverse quantization processing corresponding to the non-uniform quantization processing on the first compressed data to obtain second data, wherein the second data and second input data are used for being input into the machine learning model, and the second input data is data input into the machine learning model after the first input data. According to the present application, non-uniform quantization is performed on data on the basis of non-uniform distribution characteristics of the data, so that the distribution characteristics of the data can be better met, and a quantization result having a smaller overall error or average error can be obtained, thereby improving the processing precision of models.

The invention discloses a large model mixing precision quantification method based on dynamic tree quantification. The method comprises the following steps: step 1, dividing input feature points into outliers and non-outliers; 2, extracting an outlier from the matrix sequence of the feature points, and extracting a row where a weight corresponding to the outlier is located; or processing and quantizing the non-outlier data used for the linear operator; non-outlier data applied to a nonlinear operator is not quantized; 3, performing matrix multiplication on the outlier data and the corresponding weight of the outlier; or, performing inverse quantization on non-outlier data stored in the linear operator in a quantized manner, and performing matrix multiplication operation; directly performing matrix multiplication on non-outlier data in the nonlinear operator; and 4, combining a non-outlier matrix multiplication operation result and an outlier matrix multiplication operation result through addition, and outputting a final result. The invention further discloses application and system of the method and a storage medium.

An image decoding method according to the present invention includes reconstructing a residual block by inverse-quantizing and inverse-transforming an entropy-decoded residual block, generating a prediction block by performing intra prediction on a current block, and reconstructing an picture by adding the reconstructed residual block to the prediction block, wherein generating the prediction block includes generating a final prediction value of a prediction target pixel included in the current block based on a first prediction value of the prediction target pixel and a final correction value calculated by performing an arithmetic right shift on a two's complementary integer representation for an initial correction value of the prediction target pixel by a binary digit of 1. Accordingly, upon image encoding / decoding, computation complexity may be reduced.

The invention provides a floating-point number quantization method and device, a floating-point number inverse quantization method and device, equipment and a storage medium, and belongs to the technical field of artificial intelligence. The method comprises: based on a compression rule, determining quantization information corresponding to an index of each floating-point number in floating-point numbers to be quantized, the quantization information corresponding to the index of each floating-point number comprising a first quantization value and / or a second quantization value, the compression rule comprises a rule that an index which indicates that quantization information with a low compression degree belongs to is greater than an index which indicates that quantization information with a high compression degree belongs to, and for each floating-point number in the floating-point numbers to be quantized, based on a first quantization value corresponding to the floating-point number, performing quantization processing on a mantissa of the floating-point number, and / or performing quantization processing on the mantissa of the floating-point number based on a second quantization value corresponding to the floating-point number. And quantizing the index of the floating-point number based on the second quantized value corresponding to the floating-point number. By adopting the method and the device, the required average error can be met during quantization.

To provide a method and device for signaling chrominance quantization parameter.SOLUTION: A method for inversely quantizing a current block of a picture, includes the steps of: receiving a bitstream; acquiring a joint chrominance component residual (JCCR) control flag from the bitstream; acquiring chrominance mapping information from the bitstream on the basis of the JCCR control flag; acquiring at least one chrominance quantization parameter (QP) offset from the bit stream on the basis of the JCCR control flag; acquiring a QP value for the current chrominance block on the basis of the acquired chrominance mapping information and the at least one acquired chrominance QP offset; and executing inverse quantization on the current chrominance block by using the determined QP value.SELECTED DRAWING: Figure 7

The invention discloses a quantization method, a medium, computer equipment and a program product. The method comprises the following steps: acquiring a difference matrix between a current parameter matrix and a pseudo quantization parameter matrix of a neural network; the pseudo quantization parameter matrix is obtained by sequentially performing quantization processing and inverse quantization processing on the current parameter matrix; projecting the difference matrix to a plurality of target feature spaces to obtain a plurality of projection difference matrixes of the difference matrix; wherein any target feature space is a feature space composed of at least one feature dimension with the importance from high to low in the original feature space, and the number of the at least one feature dimension is smaller than the total number of the feature dimensions in the original feature space; determining a target projection difference value matrix from the plurality of projection difference value matrixes through grid search, and compensating the pseudo quantization parameter matrix based on the target projection difference value matrix to obtain a compensation parameter matrix; and performing quantization processing on the compensation parameter matrix.

The data processing method is applied to the field of artificial intelligence and comprises the steps that first data are obtained, and the first data are obtained through a machine learning model according to first input data; performing non-uniform quantization processing on the first data to obtain first compressed data, and storing the first compressed data to a memory; and reading the first compressed data from the memory, and performing inverse quantization processing corresponding to the non-uniform quantization processing on the first compressed data to obtain second data, the second data and second input data being used for being input into the machine learning model, the second input data is data input into the machine learning model after the first input data. According to the method, non-uniform quantization is carried out on the data based on the non-uniform distribution characteristics of the data, the distribution characteristics of the data can be better met, the quantization result with smaller overall error or average error is obtained, and thus the processing precision of the model is improved.

The present application provides an enhanced layer encoding and decoding method and apparatus. The method includes: the encoder obtains a reconstructed block of the base layer of the image block to be encoded. The difference is calculated between the corresponding pixel points of the image block to be encoded and the reconstructed block of the base layer to obtain a residual block of the enhanced layer of the image block to be encoded. The transformation block partitioning method of the residual block of the enhanced layer is determined. The residual block of the enhanced layer is transformed according to the transformation block partitioning method. The decoder obtains the bitstream of the residual block of the enhanced layer of the image block to be decoded. Entropy decoding, inverse quantization, and inverse transformation are performed on the bitstream of the residual block of the enhanced layer to obtain a reconstructed residual block of the enhanced layer. The reconstructed block of the base layer of the image block to be decoded is obtained. The corresponding pixel points of the reconstructed residual block of the enhanced layer and the reconstructed block of the base layer are summed to obtain a reconstructed block of the enhanced layer. The present application can reduce the processing flow of the encoder and improve the encoding efficiency of the encoder. By adopting an adaptive TU partitioning method, the compression efficiency of the residual block can be more effectively improved.

Embodiments of the present application provide a model training method and apparatus, and a device and a storage medium, and relate to the technical field of motion capture. The method includes: training a motion capture model according to a preset training set, wherein the motion capture model includes a time series prediction unit; performing, through a quantization node in the time series prediction unit, a quantization operation and an inverse quantization operation in sequence on model data passing through the quantization node; and adjusting a weight parameter of the time series prediction unit according to an update on a gradient of the time series prediction unit until the motion capture model converges, wherein the weight parameter includes a weight scaling parameter and a weight direction.

In one embodiment, an image decoding method for quantization shift is provided. The image decoding method may comprise a step of acquiring, from a bitstream, information indicating a quantization parameter for a current block. The image decoding method may comprise a step of acquiring a transform coefficient of the current block by performing inverse quantization on the basis of the information indicating the quantization parameter. The image decoding method may comprise a step of determining one quantization offset among a plurality of quantization offsets on the basis of at least one of a prediction mode of the current block, a frequency band of the current block, and a quantization parameter of the current block. The image decoding method may comprise a step of changing a transform coefficient by using the determined quantization offset.

The invention relates to a task-oriented semantic communication-oriented adaptive source channel coding system, which belongs to the field of wireless communication, and comprises a sending end for extracting a semantic feature map for evaluation from input data through a feature extractor initialized based on a pre-training large model, executing nested quantization and arithmetic coding based on a calculated importance score, and sending the semantic feature map to a task-oriented semantic communication-oriented adaptive source channel coding system; generating a scalable semantic bit stream with importance levels; the physical layer is used for carrying out reordering and grouping according to importance levels of bit streams, a channel encoder is combined with an interleaver to realize unequal error protection of semantic matching, and encoded bits are mapped to sign bits with different fault-tolerant capabilities in a constellation diagram through division modulation; and a receiving end: recovering the semantic feature map through arithmetic decoding and nested inverse quantization. According to the invention, on the premise of keeping compatibility with a digital communication protocol stack, through distortion-driven joint rate self-adaption, the transmission effectiveness for machine tasks and human vision is obviously improved under a limited bandwidth.

The embodiment of the invention discloses an image decoding method, an image coding method, electronic equipment and a readable storage medium. The decoding method comprises the following steps: during image decoding, performing first decoding sub-stage processing on a to-be-decoded image code stream to obtain pixel residual information; wherein the first decoding sub-stage comprises the steps of code stream acquisition and entropy decoding of an image to be decoded; caching the pixel residual error information to a shared macro block information buffer; when an image is decoded, acquiring residual information of pixels to be processed in the macro block information buffer by adopting an asynchronous independent clock; performing second decoding sub-stage processing on the to-be-processed pixel residual information to obtain image decoding information; wherein the second decoding sub-stage comprises inverse quantization, inverse discrete cosine transform and pixel code stream write-back. By decoupling the decoding process, asynchronous parallel processing of the decoding process can be realized, the image decoding efficiency is improved, and the memory utilization rate of equipment is improved.