146 results about "Uniform quantization" patented technology

A method for adaptive quantization of video frames based on bit rate prediction that includes increasing quantization in sectors of a video frame where coding artifacts would be less noticeable to the human visual system and decreasing quantization in sectors where coding artifacts would be more noticeable to the human visual system. In one embodiment the method reverts to uniform quantization for video frames in which adaptive quantization would require extra bits.

A method, system, and program product for quantizing discrete cosine transform coefficients, e.g., for MPEG compression, with minimal bit rate overhead and without using a quantization matrix. This is done by scaling a uniform quantization parameter for the entire discrete cosine transform block, defining a variety of thresholds for the quantization of discrete cosine transform coefficients below which the corresponding coefficient will be quantized to zero, and setting different normative reconstructed values for coefficients that have not been quantized to zero as the decoder will still use the original, unmodified reconstructed values as long as the corresponding coefficient is not zero.

The present invention discloses a frequency offset detecting apparatus and method for use in a digital coherent optical receiver. The digital coherent optical receiver comprises: a front-end processing section for generating a digital baseband electric signal; an equalizer for performing equalized filtering on the digital baseband electric signal; and the frequency offset detecting apparatus for detecting frequency offset contained in the digital baseband electric signal or frequency offset contained in a signal outputted by the equalizer; wherein the frequency offset detecting apparatus comprises an argument difference obtaining unit, a first subtracter, a second subtracter, a quantizer and an averager, of which the argument difference obtaining unit obtains an argument difference of adjacent symbols in a signal inputted therein; the first subtracter subtracts an output of the averager from the argument difference obtained by the argument difference obtaining unit; the quantizer performs uniform quantization with predetermined intervals on an output of the first subtracter; the second subtracter subtracts an output of the quantizer from the argument difference obtained by the argument difference obtaining unit; and the averager averages an output of the second subtracter.

A non-iterative and computationally efficient bit allocation technique for perceptual audio coders employing uniform quantization schemes. This is achieved by computing a target MNR for all critical bands in a frame using a target bit rate and associated SMRs. Associated SNRs are then computed for the critical bands using the computed target MNR and the associated SMRs. Bits are then allocated to the critical bands based on the computed associated SNRs.

The invention relates to a method for designing a code book precoded by a multi-input and multi-output system. The method comprises the following steps: according to feedback overhead of the code book, setting uniform quantization reference; according to the uniform quantization reference, constructing a first layer of the code book; calculating quantization error phi of the first layer of the code book; and using the quantization error phi as input for further quantization so as to obtain a second layer of the code book. Through the treatment of the technical proposal, the method can homogenize the quantization error caused by the limitation of the feedback overhead, thereby obtaining diversity using gain of the code book. In addition, in the presence of redundant information of a channel coding module, the advantage of the method is more obvious under the condition of low coding speed.

The present invention discloses a low-bit efficient deep convolutional neural network hardware acceleration design method based on logarithm quantization. The method comprises the following steps: S1:implementing non-uniform fixed-point quantization of low-bit high-precision based on the logarithmic domain, and using multiple quantization codebooks to quantize the full-precision pre-trained neural network model; and S2: controlling the qualified range by introducing the offset shift parameter, and in the case of extremely low bit non-uniform quantization, adaptively searching the algorithm ofthe optimal quantization strategy to compensate for the quantization error. The present invention also discloses a one-dimensional and two-dimensional pulsation arrays processing module and system byusing the method. According to the technical scheme of the present invention, hardware complexity and power consumption can be effectively reduced.

The invention discloses a Fourier graph transformation-based point cloud intraframe coding method and apparatus, and belongs to the field of point cloud digital signal processing. The method comprisesthe steps of performing voxelization on an original three-dimensional point cloud to obtain multiple point cloud voxels; clustering the obtained point cloud voxels to obtain multiple point cloud voxel sets; performing main direction weight-based Fourier graph transformation on the point cloud voxel sets; and performing uniform quantization and arithmetic coding on the transformed point cloud voxel sets to generate corresponding code streams. According to the method, the point cloud voxel sets obtained by clustering are subjected to independent graph construction, so that the graph construction complexity is lowered; the point cloud voxel sets are subjected to independent coding, so that point cloud distribution in each class is more uniform and compact; and a local similarity characteristic is fully utilized, so that the correlation between points is expressed more fully and the influence of unrelated factors such as noises and the like is reduced.

The invention discloses an all optical quantizing encoder based on nonlinear polarization rotation effects in semiconductor optical amplifiers (SOA), which belongs to the technical field of optical communication networks and mainly comprises a power allocation unit and a quantizing encoding unit based on the nonlinear polarization rotation effects in the SOAs, wherein the quantizing encoding unit mainly comprises an initial quantizing encoding module and a gain dynamic compensation module; the action of the initial quantizing encoding module is to carry out initial quantizing encoding on sampled optical pulses; and the action of the gain dynamic compensation module is to carry out gain dynamic compensation on the pulses on which the initial quantizing encoding is finished. The all optical quantizing encoder based on the nonlinear polarization rotation effects in the SOAs is capable of solving the problems that the quantizing encoder based on cross gain modulation in the SOAs is not capable of realizing high-resolution encoding and uniform quantization, and the number of the required SOAs is excessive, thereby, the all optical quantizing encoder based on the nonlinear polarization rotation effects in the SOAs has the advantages that the resolution is high, the quantization is uniform, and the number of the required SOAs is less, and simultaneously, the quantizing encoder also has the advantages of low energy consumption and convenience in photon integration and is a key technology of realizing all optical networks in future.

The invention discloses a point cloud geometric lossy compression method based on voxel convolution. Compression and decompression are carried out by carrying out convolution and deconvolution on voxels of a point cloud through a training model. Firstly, voxelization is carried out on point cloud data, a certain grid size is selected for voxels of point cloud to carry out 3D convolution operationto obtain feature data with smaller shapes and sizes, quantization processing is carried out on the feature data after convolution, uniform quantization noise is added during model training to improvethe generalization of the model, and the quantized data is compressed. During decompression, deconvolution is performed on the quantized feature data to obtain feature data of which the size is consistent with the shape and size of the initial point cloud voxel, normalizing the feature data, and judging whether each voxel unit is empty or not through a threshold value to obtain decompressed pointcloud data. During model training, focus loss is used as distortion loss to reduce the influence of too many voxel hollow samples on the model. According to the method, geometric compression can be efficiently carried out on the point cloud data, and the distortion rate after restoration is reduced.

An enhanced graph transformation-based point cloud attribute compression method. For point cloud attribute information, a point cloud is first subjected to airspace division by using a K-dimension (KD) tree; a new graph transformation processing method in combination with spectral analysis is provided; the point cloud is then subjected to spectral clustering on graphs in coded blocks of the point cloud; expansion is performed on the basis of existing graph transformation to implement a local graph transformation scheme; enhanced graph transformation with two transformation modes is formed; the compression performance of graph transformation is improved. The method comprises: performing color space transformation of point cloud attributes; dividing the point cloud by using the KD tree to obtain the coded blocks; performing spectral clustering-based enhanced graph transformation; performing transformation mode decision; and performing uniform quantization and entropy coding. Provided is a new spectral analysis-based enhanced graph transformation scheme, wherein two transformation modes are comprised, and the optimal mode is selected by the mode decision; after the point cloud is divided with the tree, a graph is created in each coded block and the graph transformation is used as transformation mode I; on this basis, graph spectral clustering is implemented; the graph is divided into two local graphs and then local graph transformation is performed to serve as transformation mode II; in the enhanced graph transformation scheme supporting the two transformation modes, the optimal mode is selected by the mode decision to achieve the optimal performance of point cloud attribute compression.

The invention discloses a method and method for generating a Gaussian white noise signal source. The method includes: using a Combined Tausworthe random number generator in an FPGA to generate long-period uniformly distributed random numbers; A combination of number generators; non-uniform quantization of long-period uniformly distributed random numbers, and segmental fitting is used to map long-period uniformly distributed random numbers to broadband white Gaussian noise; a series of filter orders are used. Multiple groups of filters composed of the same number of filter coefficients convert the broadband white Gaussian noise into an intermediate frequency noise source with adjustable parameters. Implementing the method and device for generating a Gaussian white noise signal source of the present invention has the following beneficial effects: hardware resources can be saved, and it is convenient for FPGA to realize the conversion from uniformly distributed random numbers to Gaussian white noise.

A method for quantization of a color histogram bin value of an image or video, and more particularly, a method for non-uniform quantization of a color histogram bin value of an image (or video) according to the frequency of color occurrence is provided. The methods effectively represent the characteristics of a color histogram of an image in comparison to conventional art, and improve the performance of image retrieval when an image (video) retrieval search is conducted.

The invention relates a quantization method of a dictionary learning-based image compression system and belongs to the image compression technology filed in multimedia communication. According to the method of the invention, zero coefficients are removed from a coefficient matrix; nonzero coefficient values are sorted; a nonzero coefficient sequence is truncated through using an estimated truncation coefficient percentage; a reserved nonzero coefficient sequence is normalized; a uniform quantization method is adopted to divide the processed nonzero coefficient sequence into equal subintervals; K-means clustering quantization is carried out independently in each subinterval; in the iteration process of the K-means clustering quantization, the mean value of all elements in each category is adopted as a new clustering center of the category; after an iteration termination condition is satisfied, all nonzero coefficients in each category are quantified into corresponding clustering center values; the PSNR (Peak Signal to Noise Ratio) of a reconstructed image is calculated and is compared with the set minimum PSNR given value of the reconstructed image; the truncation coefficient percentage is adjusted; and the above operation is repeated until the calculated PSNR value of the reconstructed image is not lower than the minimum PSNR given value of the reconstructed image. Compared with a quantization method according to which uniform quantization or K-means clustering quantization is used independently, the dictionary learning-based quantization method of the image compression system is advantageous in optimal quantification performance.

Methods and apparatus relating to online activation compression with K-means are described. In one embodiment, logic (e.g., in a processor) compresses one or more activation functions for a convolutional network based on non-uniform quantization. The non-uniform quantization for each layer of the convolutional network is performed offline, and an activation function for a specific layer of the convolutional network is quantized during runtime. Other embodiments are also disclosed and claimed.

The invention provides a near-lossless real-time compression method for a satellite-borne hyper-spectral interference fringe image. The method obtains the correlative characteristic of the interference fringe image by a spectral decorrelation and entropy coding method; according to the correlative characteristic of the image, the spectral decorrelation treatment is carried out on the interference fringe with a large optical path difference so as to obtain the spectral decorrelation treatment difference image; subsequently, according to the data distribution probability of the difference image, a quantization coding list suitable for the image is designed so as to carry out the lossless quantization on the value with the appearance probability more than or equal to 60% and carry out loss non-uniform quantization on the value with the appearance probability less than 60%; and finally a DPCM compression with a low compression ratio is adopted. The method has the advantages that the compressed and decompressed interference fringe image has small distortion, the compression and decompression treatment speed is quick and the real-time performance is good.

The invention provides a 5bit quantization method applied to LDPC decoding, and relates to a decoder for a communication channel. The method comprises the following steps of: based on an MS-offset algorithm, designing an information quantization expression method in a decoding process of an LDPC decoder, wherein the information quantization expression method is a corresponding conversion method mainly applied between a 5bit non-uniform quantization value and a 6bit uniform quantization value; and performing hardware design corresponding to the conversion method, thus ensuring that quantization decoding performance nearly in accordance with floating point decoding can be realized with few quantization bit number. Because the 5bit non-uniform quantization value is transmitted between a variable point calculation unit and a checkpoint calculation unit, the wiring complexity and the information memory overhead are reduced; and in the variable point calculation unit, the 5bit non-uniform quantization value is converted into the 6bit uniform quantization value and then operation is performed so as to enlarge the expression range of the quantization information and improve the decoding performance.

The invention provides a K-means non-uniform quantization algorithm for filter bank multi-carrier modulation optical communication system. The K-means non-uniform quantization algorithm comprises thefollowing steps: step 1, inputting a data sequence sent by a filter bank multi-carrier modulation transmitter into a data preprocessing module, and taking a generated vector as the input of a K-meansquantization module; 2, selecting K vectors as an initial quantization order, calculating the distance between each vector and a quantization order vector, clustering each vector into a set of quantization order vectors with the minimum distance according to the distance, and calculating a quantization interval and a central vector of each quantization interval; step 3, taking the center vector generated in the step 2 as an initial quantization order of distance calculation, and iterating the step 2 until convergence to obtain K quantization intervals and center vectors; and 4, outputting thequantization interval and the central vector obtained in the step 3 to obtain a quantization result. The K-means non-uniform quantization algorithm reduces the quantization error of the system, improves the bit error rate performance of the system, and saves the cost.

The invention provides a deep learning feature compression and decompression method and system, and a terminal, and the method comprises the steps: carrying out the compact space transformation of an input original feature through a coding end, and obtaining the compact feature expression of the original feature; calculating an importance coefficient of each channel in the features by using compact feature expression, and performing quantization parameter adaptive distribution of each channel; performing non-uniform quantization on different channels based on the distributed quantization parameters to obtain a quantized multi-channel feature map; and performing feature coding on the quantized multi-channel feature map to complete compression. The decoding end carries out decoding to obtain an inverse quantized multi-channel feature map; self-adaptive performance compensation is carried out on different quantization levels of the inversely quantized multi-channel feature map; and original feature reconstruction is performed on the compensated multi-channel feature map to complete feature decompression. According to the method, the coding end adaptively allocates the quantization parameters based on the image content, and the decoding end adaptively performs performance compensation based on the quantization level, so that the performance of non-uniform quantization on feature compression is improved.

The invention discloses a nonlinear modeling method of a wide-band transmitter based on dynamic multi-core bandwidth generalized regression neural network algorithm. The method comprises the followingsteps: constructing a test platform, measuring and recording nonlinear characteristics of a full working frequency band of the wide-band transmitter, and collecting input and output amplitudes and aphase value of a test signal, and corresponding carrier frequency; performing segment processing on a signal sample according to the carrier frequency, quantifying the amplitude of an input signal byusing an amplitude non-uniform quantization technology, and constructing a dynamic nonlinear fingerprint sample data set by using the dynamic memory fingerprint sample technology; and training a generalized regression neural network by using the dynamic nonlinear fingerprint sample data set, obtaining a dynamic multi-core bandwidth generalized regression neural network model in combination with the optimization algorithm, and achieving full-band nonlinear characteristic modeling of the wide-band transmitter.