194results about How to "Reduce quantization error" patented technology

The invention provides a depth neural network based vector quantization system and method, comprising a normalization preprocessing module for normalizing original data through normalized data and outputting preprocessed data after normalization; a vector quantization and coding module for receiving the preprocessed data and the codebook and carrying out vector quantization coding to the preprocessed data through the codebook and outputting the coded data; a neural network inverse quantization module for performing the decoding of the inverse quantization to the coded data through a depth neural network and outputting the decoded data; a processing module after inverse normalization for performing an inverse normalization process to the decoded data through the normalized data and outputting the restored original data after the inverse normalization; and a neural network training module for carrying out trainings to the neural network through the pre-processed training data and coded training data after normalization processing and outputting the neural network to the neural network inverse quantization module. The system and the method of the invention can effectively solve the problem that the quantization error is large in high dimension signal vector quantization.

The invention discloses a DSAH (discrete semantic alignment hashing) method based on semantic alignment for cross-modal retrieval. In the training process, a heterogeneous gap is reduced by the aid ofimage attributes and modal alignment semantic information. In order to reduce memory overhead and training time, a latent semantic space is learned by synergistic filtering, and the internal relationbetween a hash code and a label is directly built. Finally, in order to decrease quantization errors, a discrete optimization method is proposed to obtain a hash function with better performances. Inthe retrieval process, samples in a testing set are mapped to a binary space by the hash function, the Hamming distance between a binary code of a query sample and a heterogeneous sample to be retrieved is calculated, and front ranked samples are returned according to the sequence from small to large. Experimental results of two representative multi-modal data sets prove superior performances ofDSAH.

The invention discloses a channel quantification method for a multi-user MIMO system based on limited feedback and a device thereof, which belongs to the technical field of wireless transmission. The method comprises the following steps: according to a channel state message, from a codebook choosing a codeword which has the smallest angle with the codebook and calculating an quantification error; choosing a codeword vector which has the smallest angle with the quantification error from the codebook; projecting a channel vector to the space formed by the two chosen codeword vectors, calculating and obtaining a compound proportion factor; quantifying the compound proportion factor by a pre-generated quantification step set. A receiver device comprises a channel state message obtaining unit, a channel vector qualification unit, a quantification message feedback unit. A sender device comprises a feedback message reconstruction unit and a pre-coding unit. The method and the device of the invention can reduce the quantification error and improve the capacity performance of the system and has the advantages of small codebook and low complexity with same rate and capacity.

The invention discloses a speech-emotion recognition method based on improved fuzzy vector quantization. The method extends the sum of fuzzy membership function from one to N so as to reduce the influence of sample wild-point on an iteration-training process to a certain extent, and adopts a clustering method based on similarity threshold and a minimum distance principle in the iteration-training process so as to avoid the problem that a clustering center is sensitive to initial values and easy to fall into local minimum values to a certain extent. Experimental results prove that the method can effectively improve the emotion recognition rate of the prior fuzzy vector quantization method.

The invention discloses a microgrid multi-inverter parallel-control method based on frequency division virtual complex impedance, which is suitable for a multi-inverter parallel-control system with a nonlinear load for a low-voltage microgrid. The microgrid multi-inverter parallel-control method comprises the following steps: performing frequency division on a line current by virtue of a band pass filter by adopting a robust droop control strategy of resistance equivalent output impedance, simultaneously multiplying a fundamental current by a virtual complex impedance containing a resistance-capacitance component, designing that each parallel connection inverter output impedance is resistive in a fundamental band; and respectively multiplying each high-frequency current by different virtual inductive impedances, indicating that a harmonic frequency band inverter output impedance is inductive so as to be limited by the robust droop control strategy of resistance equivalent output impedance, so that the influence of a high-frequency harmonic current to inverter parallel-control is avoided. According to the microgrid multi-inverter parallel-control method, the defect of a conventional method is overcome, and good control effect is achieved in the multi-inverter parallel-control system with the nonlinear load for the low-voltage microgrid.

Disclosed are a code book creation method and device and an initial code book creation method, for a wireless communications system. The code book creation method includes: a step in which the amplitude or power for channel status information for channels between a plurality of transmission antennas and a reception antenna is obtained; and a step in which elements of a codeword for an initial code book are dynamically mapped to the plurality of transmission antennas and a first code book is generated, on the basis of the amplitude or the power. The initial code book creation method includes: a step in which prescribed elements of the codeword are set to have a prescribed phase resolution; and a step in which the other elements of the codeword are set so as to be arranged in order of gradually increasing or gradually decreasing phase resolution. The order of arrangement between elements having the same phase resolution is arbitrary and the phase resolution is the number of phases possible for all elements in the same position for all codewords in the initial code book. As a result, the accuracy of the code book and the performance of the system are improved.

The invention discloses a method and a device for detecting speech keywords as well as a corresponding retrieval method and a system thereof. The system for detecting the speech keywords comprises a VQ code library, a state characteristic library, a state sequence generating unit and a matching unit, wherein the VQ code library stores VQ codes based on a preset acoustical model; the state characteristic library stores a VQ state sequence matrix of an audio file, and the VQ state sequence matrix comprises VQ code numbers and corresponding acoustical probability values of various audio frames of the audio file; the state sequence generating unit is based on the VQ codes in the VQ code library and converts an input keyword into a corresponding VQ state sequence; and the matching unit utilizes the converted VQ state sequence to reduce the VQ state sequence matrix in the state characteristic library and utilizes the VQ state sequence to search the speech keyword in the reduced VQ state sequence matrix. The invention technology can be utilized to rapidly search the speech keywords in the audio file and rapidly find the audio file containing the speech keyword in an audio file library.

The invention relates to a nonuniform quantizing coding method used for decoding an LDPC code and the application of the method in a decoder. The method comprises the steps that the quantification range of messages L(rji), L(qij) and L(Qi) is determined according to information distribution of an initial message L(Pi), a coding mode is adopted for performing nonuniform quantizing on the messages L(rji), L(qij) and L(Qi), quantized intervals of positive number portions and negative number portions are respectively divided into m sections, coding is a codon composed of 1-bit polar codes, k-bit segment codes and (n-1-k)-bit in-segment codes, the polar codes mark positive numbers and negative numbers, the segment codes mark different segments, the segment internal codes show specific size of the data in the segments, each segment is divided into 2<g> small segments, and the small segments are the quantized intervals of the corresponding segments. According to the method, consumption of hardware storage resources, operation processing units and locating and wiring resources is reduced, quantization errors of small-value data are reduced, and the decoding performance is improved.

The invention provides an unsupervised cross-modal hash retrieval method and system based on virtual label regression, and the method comprises the steps: feature representation and hash function learning are integrated into a unified depth frame, and a shared hash code is learned through the cooperative matrix decomposition of multi-modal depth features, so as to guarantee that a plurality of modes share the same semanteme; on the basis, the concept of a virtual label is introduced, the virtual label is learned through non-negative spectrum analysis, and meanwhile, the learned virtual labelis returned to the hash code, so that the semantic consistency between the hash code and the virtual label is ensured; in the framework, collaborative matrix decomposition of the depth features and learning and regression of the virtual tags are beneficial to learning of depth feature representation and hash functions, and improved depth feature representation and hash models are beneficial to collaborative matrix decomposition and learning and regression of the virtual tags, and the collaborative matrix decomposition and the virtual tags promote each other; and meanwhile, through a new discrete optimization strategy, the deep hash function and the hash code are directly updated, the quantization error of a relaxation strategy in an existing method is effectively reduced, and the cross-modal retrieval performance is improved.

The invention discloses a distributed sensing system based on the differential amplification technology. The system is composed of an optical path detection part and a circuit demodulation part. In an optical path, two different frequencies of microwave signals are modulated to signal light, the frequency of one path of microwave signals is within the Brillouin frequency-shift frequency, and the frequency of the other path of microwave signals is slightly lower than the Brillouin frequency-shift frequency. and in a circuit, a power divider separates the two paths of signals, the separated two paths of signals are filtered and demodulated respectively, and the demodulated signals are amplified in the differential amplification manner. The signals collected by a data collection card are amplified within the specific amplitude interval by the differential amplification structure, and the problems that the range of the data collection card of a host in the distributed Brillouin time-domain analysis sensing system is limited and that the data collection precision cannot be satisfied and further the measurement accuracy of the system is reduced when the DC component is very large are solved. Devices in the system work stably, the point of temperature or stress change and the temperature or stress value can be accurately measured in multiple times of measurement.

The invention discloses a zeroshot hash image retrieval method based on supervision transfer, belonging to the technical field of image hash. Modeling is performed on a label with training data by using an existing natural language processing model, and a label space is formed; by means of a relationship between potential storage labels in the new label space, mapping from an image feature space to the label space is trained; on the basis, the mapping relationship is reflected on a hash code. Strict requirements on training samples are avoided, the application scope is wide, and especially in a large database, when categories are more and training samples cannot be found for each category, the image retrieval accuracy rate of the categories can be greatly improved by the method.

The invention discloses an IRS-based large-scale MIMO cascaded channel estimation scheme under a hybrid low-precision architecture, which comprises the following steps: firstly, a base station end performing uniform division and phase configuration of units on a large-scale reflective array plane by sending a downlink configuration information signaling, and selecting a central unit of each sub-array plane to perform channel estimation; secondly, designing an isotropic analog receiver which adopts a full-connection phase shifter network, and enabling phases of all phase shifters to be uniformly distributed; then, the base station obtaining sparse information of an angle domain millimeter wave channel by utilizing a compressed sensing technology, so that a digital estimator is optimally designed, and angle domain channel estimation is obtained; and finally, through sparse recovery and angle domain inverse transformation, obtaining the cascade channel estimation of the sub-array surface,and further deducing to obtain the cascade channel estimation of all the IRS array surfaces. According to the method, the performance of channel estimation can be remarkably improved, and the amplification effect of the IRS on quantization noise is effectively inhibited.

The invention discloses a method for establishing a passive target measurement model based on a wireless sensor network. The method is characterized by comprising the following steps: 1. computing the position relations of the target position relative to wireless sensor nodes, thus obtaining an ellipse with the wireless sensor nodes i and j as the focuses; and 2. determining attenuation of the wireless signal from the wireless sensor node i to the wireless sensor node j. The method has the following advantages and positive effects: the passive target measurement model is established by judging the position relations of the pre-estimated target position relative to the wireless sensor nodes and computing attenuation of the strength of the signal received in the position, thus overcoming the restriction to pixeling treatment of the monitored area, greatly adding the application context in target positioning and tracking, avoiding pre-pixeling treatment from being carried out on the monitored area and reducing artificially introduced quantization errors, so better positioning and tracking precision can be achieved.

The invention provides a combined optimizing method for the sizes of the multiple vector and multiple codebooks on the basis of a superframe mode. The invention belongs to the technical field of the compressed encoding of sound and is characterized in that: the method firstly train the quantified codebook according to the quantified bit number distributed by bandpass surd and sonant parameters, and define the superframe mode according to the quantified codebook; jointly allocate the codebook sizes of the three under each superframe mode according to quantified bit numbers distributed by the keynote cycle parameters, line spectrum frequency parameters and the short-time frame energy parameters; train the corresponding quantified codebook according to the codebook sizes of each parameter under different superframe modes, then carry out vector quantification to the three parameter vectors. The method can improve the quantification accuracy of the parameters, reduce the error of the synthesized speech and enhance sound articulation and is most suitable for the encoding of low-rate speech.

The invention discloses a cross-modal retrieval method and device based on deep adversarial discrete hash learning, and the method comprises the steps: forming an image network through three full connection layers, forming a text network through four full connection layers, and obtaining a final hash value through the activation of a tanh function. Through adversarial training of the two networks,it can be ensured that feature expressions of the two modes tend to be consistent; semantic similarity is guaranteed through a cosine triple loss function to be weighted, meanwhile, relevancy ordering of similar samples is guaranteed, discreteness of hash values is maintained in the optimization process, quantization errors are reduced, and finally two robust hash functions for keeping semantic similarity, reducing heterogeneous gaps and being small in accumulative error are obtained. According to the method, the hash function is learned by utilizing adversarial training and keeping semanticsimilarity and relevancy ranking, so that the retrieval precision is improved, and the obtained hash learning method has stronger semantic learning capability.

The invention relates to a high-speed high-precision flow line structure ADC. A circuit comprises N one-bit analog-to-digital conversion sub-modules which are connected in series step by step; each one-bit analog-to-digital conversion sub-module comprises a comparator, an alternative analog data selector and a subtracter with gain fixed at 2; the one-bit analog-to-digital conversion sub-module of each stage compares an input signal V and a reference signal D of the module, outputs a comparison result, namely the divided quotient (under binary sense) of the two signals, and outputs 2 times of the divided remainder (under binary sense) of the two signals as the input of next stage; and N conversion data are arranged according to a weight order to acquire the output of the whole analog-to-digital converter. The analog-to-digital converter is provided with a 'rounding' offset circuit which is used for improving the precision of the lowest bit. Therefore, the precision can reach 1/2 of resolution, namely LSB/2 (Vr/2<N+1>).

Disclosed is an in-band full duplex transceiver. The in-band full duplex transceiver may include a transmitter, a hybrid transformer, and an impedance matching unit. The hybrid transformer may include a first end connected to an antenna, may output a transmitting signal to the antenna, and may output a received signal provided through the antenna to the receiver. The impedance matching unit may include balance networks connected to a second end of the hybrid transformer and matching impedance of the antenna.

The invention discloses a method for designing double-codebook difference in a LTE-A terminal feedback system. The correlation of the time and the space of a communication channel is fully utilized, a W1 codebook is rotated regularly according to a feedback precoding matrix indicator to better track the main direction characteristics of the communication channel, meanwhile, a W2*W1 codebook is zoomed to reduce the quantization error of the communication channel, and W = W2*W1 is used for expressing the communication channel information after quantization; a receiving end calculates a precoding matrix by using a zero-forcing beam forming method according to the feedback quantization communication information. According to the method, interference among multiple users is removed while low feedback overheads are guaranteed, and the overall performance of MIMO of the users is improved. Meanwhile, the precoding matrix is calculated according to the feedback quantization communication channel information, the interference among the users is removed, the performance of an MIMO system of the users is improved, and the performance gain is about 1-2dB when the signal to noise ratio is high.

A gamma correction circuit for correcting a digital video signal, the circuit comprising first (6) and second (8) lookup tables for storing discrete output intensity data and the associated slope data of a non-linear transfer function, respectively, for each of the discrete input video signal intensities, an adder (10) having a first input connected to the output of the first lookup table, a multiplier (12) having a first input connected to the output of the second look-up table (8), characterized by a quantizer (4) for providing the most significant bits of the incoming video signal to address the first (6) and second (8) lookup tables and to transfer the corresponding output intensity data to the adder (10) and the associated slope data to the multiplier (12), the quantizer (4) transmitting the remaining least significant bits of the input video signal to the second input of the multiplier (12), the multiplier (12) multiplying the slope data with the remaining least significant bits and feeding the multiplication result to the second input of the adder (10), the adder (10) adding the output intensity data and the multiplication result to generate a corrected video signal.

The invention discloses a channel state information compression method and a channel state information compression device, which belong to the field of wireless communication. The method comprises the following steps of: searching for a replaced code word from a Huffman coding codebook, wherein the probability of occurrence of the replaced code word is less than a preset threshold value; calculating Euclidean distances between the replaced code word and each of the other code words in the Huffman coding codebook, selecting a replacement code word according to calculation results, and replacing the replaced code word by utilizing the replacement code word, thereby obtaining a processed Huffman coding codebook; and coding the channel state information quantification result of a current moment according to the processed Huffman coding codebook and the channel state information quantification result of the previous moment. In the method and the device, the code words with the probabilities of occurrence less than the preset threshold value are replaced by the code words at the shortest Euclidean distance from the code words with the probabilities of occurrence less than the preset threshold value, so that the use of the code words with the probabilities of occurrence less than the preset threshold value are avoided in the coding of the channel state information quantification result, a coding length is greatly shortened, the coding efficiency is improved, the coding is dynamically related with time, and the dynamic adjustability of the coding is enhanced.