64 results about "Binary neural network" patented technology

The invention provides a face image privacy protection recognition method based on optical coding and reverse feature enhancement, and relates to the technical field of computer vision recognition, and the method comprises the steps: S1, obtaining a face data set, carrying out the data enhancement, and constructing a face enhancement data set; s2, constructing a binary neural network convolutional coding layer based on straight-through estimator quantization, and performing optical coding on the face enhancement data set; s3, constructing a reverse feature enhancement neural network model based on CNN and Transform double-channel multiplexing, and carrying out feature dimension increase; s4, performing joint optimization on the neural network coding layer and the dual-channel multiplexing reverse feature enhancement neural network model by using mixed constraints; and S5, constructing a face embedding database, and comparing the high-dimensional coding value of the single-pixel detection value with features in the face embedding database to realize face recognition. According to the invention, through joint optimization of the neural network binarization convolution coding layer used for optical coding of the spatial light modulator and the dual-channel multiplexing reverse feature enhancement model, the recognition precision is improved.

The invention relates to a training method and device of a binary neural network model and computer equipment. The method comprises the steps that enhanced thermometer coding is introduced into an input layer of a small sample image recognition model, and a binary neural network model is designed by adopting a batch-free normalized convolutional layer; inputting a to-be-predicted picture to the binarization neural network model to extract to-be-predicted features, performing first-stage training on the to-be-predicted features by using the original activation function and the weight, and performing second-stage training on the predicted picture input to the trained binarization neural network model by using the binarized original activation function and the weight, and obtaining an optimal binary neural network model. And deploying the optimal binary neural network model into a C language reasoning framework, completing system module development, and obtaining edge device adaptation information. And designing a register transfer level process of the target edge device. By adopting the method, the efficient and lightweight performance of deep learning of the resource-limited edge device can be improved.

An all-binary neural network system and method for processing and analyzing multi-source time series data is disclosed. The system employs a shared codebook to encode input streams into binary codewords, which are then processed through a series of binary convolutional layers, binary LSTM layers, and binary fully connected layers. The system maintains binary representations throughout, enabling efficient computation and reduced memory requirements while effectively capturing temporal and inter-source relationships in the data.

The invention discloses an electric energy quality disturbance high-precision classification method of a hybrid binary neural network, and the method comprises a full-precision convolution layer, a binary convolution layer, and a binary full-connection layer used for classification. Comprising the following steps: retaining original input information of a power quality disturbance signal to the greatest extent through a full-precision convolutional layer; adopting a binary convolution layer to reduce the size of the model, and performing feature extraction on the data; through a pooling layer and a batch normalization layer, the training speed in the feature extraction process is increased, and overfitting is reduced; and finally, classifying the extracted features by using a binary full-connection layer as a classifier and reducing the size of the model at the same time. And outputting a corresponding category by a Softmax function to realize power quality disturbance classification. According to the method, a small number of full-precision convolutional layers are added in the binarized convolutional neural network, so that sample features are reserved to the maximum extent, the model size is reduced, the classification speed is improved under the condition that the efficiency is controllable, and high-precision classification of the power quality disturbance is achieved.

The invention provides a quasi-e-system neural network model, a construction method, a data processing method and an electronic device, the quasi-e-system neural network model comprises an input layer, a hidden layer and an output layer, the input layer is used for receiving data; the hidden layer comprises neurons, the neurons comprise at least one sub-neuron, the sub-neuron can represent any one of a binary function g (x) and a ternary function t (x), and an activation function F (x) of the neurons represents linear weighted fitting of the binary function g (x) and the ternary function t (x). According to the scheme provided by the invention, by adopting the combination of the binary function g (x) and the ternary function t (x), the precision and the energy consumption of the model are balanced on the basis of keeping the high compression ratio and the high calculation efficiency of the binary neural network.

Apparatuses and methods can be related to implementing a binary neural network in memory. A binary neural network can be implemented utilizing a resistive memory array. The memory array can comprise programmable memory cells that can be programed and used to store weights of the binary neural network and perform operations consistent with the binary neural network. The weights of the binary neural network can correspond to non-zero values.

The application discloses a permanent magnet synchronous motor stator fault diagnosis method and system based on a human learning optimization algorithm, and the method comprises the following steps: obtaining a three-phase current signal of a stator of a permanent magnet synchronous motor, converting the signal into a two-dimensional time-frequency diagram after being decomposed by a variational mode and being demodulated by a Hilbert envelope; constructing a binary neural network for processing the two-dimensional time-frequency diagram and determining a search space; generating a candidate architecture by global search using a first human learning optimization algorithm; for each candidate architecture, fine-tuning the classification layer parameters thereof by proportional evolution using a second human learning optimization algorithm, taking the classification performance of a verification set as a fitness value, and iteratively obtaining an optimal model; inputting a current signal to be diagnosed into the model after the same preprocessing, and outputting a stator fault diagnosis result. Compared with the prior art, the application realizes low-overhead accurate diagnosis of weak faults of a permanent magnet synchronous motor stator by physical information driven preprocessing and double-layer human learning optimization.

The invention discloses a binary neural network pruning compression and deployment optimization system and method, and aims to solve the problems of low PopCount calculation efficiency, non-fusion of pooling and convolution, unstable pruning training and the like in the prior art. The system comprises a regularization item module with an attenuation factor during training, and a binary convolution runtime pruning module, a PopCount lookup table module and a binary convolution and binary maximum pooling fusion calculation module during deployment. The method comprises the following steps: 1) adjusting regularization intensity by adopting a dynamic attenuation factor, and improving pruning training stability; the method comprises the steps of (1) obtaining a binary neural network, (2) terminating redundancy calculation in advance through a dual-condition pruning mechanism to optimize reasoning efficiency, (3) achieving efficient PopCount calculation based on a segmented lookup table, and (4) fusing binary convolution and pooling operation to reduce boundary checks.According to the method, the calculation efficiency and deployment adaptability of the binary neural network on end-side equipment are remarkably improved, and the method is suitable for embedded scenes with limited resources.

A framework is presented that provides a shift in the conceptual and practical realization of privacy-preserving interference on deep neural networks. The framework leverages the concept of the binary neural networks (BNNs) in conjunction with the garbled circuits protocol. In BNNs, the weights and activations are restricted to binary (e.g., ±1) values, substituting the costly multiplications with simple XNOR operations during the inference phase. The XNOR operation is known to be free in the GC protocol; therefore, performing oblivious inference on BNNs using GC results in the removal of costly multiplications. The approach consistent with implementations of the current subject matter provides for oblivious inference on the standard DL benchmarks being performed with minimal, if any, decrease in the prediction accuracy.

A binary neural network (BNN) without floating-point layers uses inter-layer input data and inter-layer output data represented in n-bit binary format, where n is less than 32. For instance, n may be 8. Each layer of a set of hidden layers is operative to perform binarized operations on respective input data and at least one of the hidden layers, and includes a trained one-bit binary quantizer with a preconfigured binarization threshold that was configured during training.

This invention proposes a lightweight remote sensing image change detection method based on a binary neural network and large-kernel strip convolution. The implementation steps are as follows: obtaining training and testing sample sets; constructing a lightweight remote sensing image change detection network model based on a binary neural network and large-kernel strip convolution and iteratively training it; and obtaining the lightweight remote sensing image change detection results. In the process of iteratively training the lightweight remote sensing image change detection network model and obtaining the change detection results, this invention uses ResNet18 networks in two branches to perform deep feature extraction on the feature maps of both time periods. This allows the weights and activations to be quantized to +1 and -1, achieving a 32-fold memory saving and a 58-fold CPU acceleration, avoiding the drawback of existing technologies with a large number of parameters, and effectively improving detection efficiency. Simultaneously, the large-kernel strip convolution network compensates for the accuracy deficiency caused by binary quantization.

The application provides a face image privacy protection recognition method based on optical coding and reverse feature enhancement, and relates to the technical field of computer vision recognition, and the method comprises the following steps: S1, acquiring a face data set, performing data enhancement, and constructing a face enhanced data set; S2, constructing a binary neural network convolution coding layer based on a straight-through estimator quantization, and performing optical coding on the face enhanced data set; S3, constructing a CNN and Transformer dual-channel multiplexing reverse feature enhancement neural network model, and performing feature dimension enhancement; S4, using a hybrid constraint, and jointly optimizing the neural network coding layer and the dual-channel multiplexing reverse feature enhancement neural network model; and S5, constructing a face embedding database, comparing a high-dimensional coding value of a single-pixel detection value with features in the face embedding database, and realizing face recognition. The application jointly optimizes the neural network binary convolution coding layer for spatial light modulator optical coding and the dual-channel multiplexing reverse feature enhancement model, so that the recognition precision is improved.

The present invention discloses a voice wake-up method and system based on a time-domain binary neural network. The method comprises the following steps: obtaining an audio file to be recognized, thereby obtaining a voice signal to be processed; extracting acoustic features from the voice signal to be processed, and performing dimensionality conversion processing on the voice signal to obtain acoustic features after dimensionality conversion processing; inputting the acoustic features after dimensionality conversion processing into a pre-trained time-domain binary neural network (TBNN) model to obtain probability outputs of keywords and non-keywords; and determining whether to wake up the system based on whether the highest probability among the probability outputs of keywords and non-keywords is a keyword. The present invention greatly reduces the number of parameters and the amount of computation of the neural network classifier, while significantly improving the wake-up speed and reducing the power consumption of the voice wake-up system.

A system and method for single cycle binary matrix multiplication in neural network computations is disclosed. The system includes a memory array storing binary weights, an input unit for activating rows based on a binary activation vector, and per-column majority sense amplifiers. The system performs binary matrix multiplication in a single cycle, enabling efficient implementation of binary neural networks. The memory array may include sections for weights and inverse weights, with corresponding activation register sections. Differential sense amplifiers may implement the majority function. The system can be applied to convolutional neural networks, using SRAM arrays for image storage and processing. Methods for determining majority votes and counting activated bits using iterative modification of the activation vector are also described.

The present invention discloses a target detection accelerator and a binary quantization training method thereof, belonging to the field of neural network accelerator quantization and hardware deployment. The multiplication module in the convolution calculation unit is replaced by a lookup table module, which reduces the amount of multiplication calculation in the convolution process, and fully utilizes the characteristics of weight binarization to reduce the redundancy in the calculation process. An efficient binary neural network accelerator hardware data flow module is designed. According to the characteristics of the binary target detection neural network model, the input data reuse method is adopted to maximize the reuse of the pre-calculated results obtained in the lookup table of the convolution pre-calculation unit, thereby minimizing the redundancy in the accelerator calculation process. In the design of the hardware data flow, the advantage of the low bit width of the weight is fully utilized to reduce the operating power consumption of the accelerator. The accelerator is binary quantized and trained with a preset data set, and the weight values ​​in the network are quantized to 1 bit, which reduces the computational complexity and storage requirements of the model.

A local activation method and system based on a binary neural network, said method including: during forward propagation, comparing the difference between a center pixel and an adjacent pixel to determine a local activation value; during forward propagation, setting an appropriate number of local activation channels and an activation direction to obtain a local activation feature map having different activation directions; during forward propagation, using weighting coefficients, which can be learned, to perform channel fusion on the output feature map after local activation and direct activation, and obtaining an output feature map containing both texture features and contour features; during backpropagation, using an asymptotic sine function to update the weights of the binary neural network. The invention is capable of effectively reducing the loss of information during binary activation, and can effectively reduce gradient mismatch during backward gradient update of a binary neural network, thereby improving the performance of the binary neural network.

The application discloses a cooking robot, an intelligent controller and a monosodium glutamate feeding method. The method comprises the following steps: feeding monosodium glutamate into a monosodium glutamate dissolving pool to form a monosodium glutamate aqueous solution; acquiring a monosodium glutamate aqueous solution image of the monosodium glutamate dissolving pool, and performing image analysis on the monosodium glutamate aqueous solution image in combination with a region clustering algorithm and a microscopic binary neural network; judging whether the monosodium glutamate inventory in the monosodium glutamate dissolving pool is up to standard based on the image analysis result; if it is judged that the monosodium glutamate inventory in the monosodium glutamate dissolving pool is not up to standard, continuously feeding monosodium glutamate into the monosodium glutamate dissolving pool until the monosodium glutamate inventory in the monosodium glutamate dissolving pool is up to standard; and conveying the monosodium glutamate aqueous solution in the monosodium glutamate dissolving pool into corresponding cookware. The monosodium glutamate aqueous solution image is quickly processed and accurately analyzed in real time by combining the region clustering algorithm and the microscopic binary neural network, the monosodium glutamate in the monosodium glutamate dissolving pool can be timely supplemented, and therefore, automatic blending and accurate feeding of monosodium glutamate in a cooking process can be realized.

A system and method for single cycle binary matrix multiplication in neural network computations is disclosed. The system includes a memory array storing binary weights, an input unit for activating rows based on a binary activation vector, and per-column majority sense amplifiers. The system performs binary matrix multiplication in a single cycle, enabling efficient implementation of binary neural networks. The memory array may include sections for weights and inverse weights, with corresponding activation register sections. Differential sense amplifiers may implement the majority function. The system can be applied to convolutional neural networks, using SRAM arrays for image storage and processing. Methods for determining majority votes and counting activated bits using iterative modification of the activation vector are also described.

The invention relates to the field of semiconductor devices, and discloses an in-memory computing accelerator and an in-memory computing system for a ferroelectric capacitor without an operational amplifier, a word line WL and a word line anti-WLB are parallel to each other, and a bit line BL is located in the vertical direction of the word line WL and the word line anti-WLB; one end of the first capacitor is connected with the bit line BL, and the other end is connected with the word line WL; one end of the second capacitor is connected with the bit line BL, and the other end of the second capacitor is connected with the word line reverse WLB. According to the calculation accelerator in the ferroelectric capacitor memory, by optimizing an array structure and a read-out mechanism, high power consumption caused by OPAMP is eliminated, meanwhile, high calculation precision and energy efficiency are kept, and the calculation accelerator is suitable for a high-efficiency reasoning task of an edge-end binary neural network BNN.

The present disclosure relates to a circuit comprising a first memory element configured to store a first data value; a second memory element configured to store a weight matrix in association with a layer of a binary neural network; and a computing circuit configured to: a) receive the first data value and the k-th row of the weight matrix; b) receive a first control signal, indicating the nature of each of the first and second read functions, each associated with two-valued arithmetic; c) generate a first vector by applying the first read function to the k-th row of the weight matrix and a second vector by applying the second read function to the data; and d) generate a k-th component of a first output vector based on the first and second vectors.

The invention relates to the technical field of currency detection, and discloses a currency detection method, system and device based on a deep binary neural network, and the key point of the technical scheme is that the method comprises the following steps: S1, obtaining an image of a currency sample and corresponding labeling data, and forming a data set, the currency sample comprises a variety of genuine currency and counterfeit currency, and the labeling data comprises true and false labels; s2, constructing a deep binary neural network model, wherein the model comprises an input layer, a plurality of binarization convolution layers, a plurality of binarization full-connection layers and an output layer which are connected in sequence; s3, training, verifying and testing the deep binary neural network model through the data set to obtain an applicable model; and S4, integrating the applicable model in currency detection equipment, and detecting the to-be-detected currency.

The present invention proposes an image classification method and system based on a binary neural network, comprising: constructing a neural network module including a feature amplification layer, a binary convolution layer, an activation layer and a segmented scaling layer, and constructing a binary neural network by stacking the neural network modules; obtaining an image with an image category label marked as training data, inputting a floating-point feature map of the training data into the feature amplification layer of the first module in the binary neural network to amplify the number of channels of the floating-point feature map to obtain an amplified feature map, converting the amplified feature map into a binary feature map and inputting it into the convolution layer to obtain a convolution feature map of the binary feature map, normalizing the convolution feature map and inputting it into the segmented scaling layer, adjusting the floating-point feature map output by the scaling factor of the segmented scaling layer, and passing the result as input to the next neural network module, and using the image category of the training data obtained by the last neural network module as the training result.

Methods and systems for computing in-dynamic random access memory (DRAM) computing include loading a first group of cells of the DRAM with input parameters and loading a second group of cells of the DRAM with inverted input parameters that are each complementary to corresponding input parameters. An offset group of cells of the DRAM is loaded with an indication of an offset voltage. An operation is performed on weights with corresponding stored input parameters or the stored inverted input parameters, and a column of the first group and the second group is activated to perform an accumulation of the operations of weights for cells in the column to store a sum. An offset voltage is generated using the indication, and an output is generated based on the comparison of the sum and the offset voltage and is stored in an output group of cells of the DRAM.

This patent relates to an asymmetric channel expansion and contraction technique (AESBNN) applied to binary convolutional layers, belonging to the field of binary neural network optimization technology in deep neural networks. This design aims to solve the problem that existing symmetric channel expansion and contraction techniques (ESBNN) cannot balance computational complexity and model accuracy according to the needs of different scenarios due to their fixed expansion / contraction ratio. The core solution is to introduce an asymmetric ratio r to first expand the number of input channels by a factor of t to increase the dynamic range of the binary convolution output. Then, the number of output channels of the convolution kernel is reduced to 1 / (t*r) of the original number to adjust computational complexity. Finally, through channel duplication and splicing layers and batch normalization (BN) layers, the final output channels are restored to be compatible with subsequent network layers. This method provides a flexible accuracy-efficiency balance strategy: when r > 1, computational complexity is prioritized to be reduced; when r < 1, the model's representational ability is prioritized to be improved. Experiments show that this technology can achieve better performance (such as PSNR) than basic binary convolutional units in tasks such as image super-resolution with lower computational cost (such as halving the computational cost when r=2), and can be flexibly integrated into various deep neural networks as a plug-and-play general module.

The invention provides a system for adjusting the flatness of a mask coiled material. The system comprises a mask coiled material horizontal linear transmission area formed by a plurality of conveying rollers, a flatness adjusting roller group, a line laser scanner and a calculation processing module, and all rollers of the flatness adjusting roller group abut against the segmented areas of the mask coiled material, and the segmented areas are overlapped to completely cover the width of the mask coiled material. And the calculation processing module continuously acquires a reflection light band deformation image formed by the line laser on the upper surface of the mask in the scanning area through the line laser scanner, and inputs the reflection light band deformation image to the pre-trained lightweight binary neural network to determine whether the flatness of the mask coiled material in the scanning area is qualified. And when the flatness of the mask coiled material is not qualified, according to the continuous cross section contour of the upper surface of the mask coiled material in the scanning area, the height of each roller of the flatness adjusting roller group is adjusted in an auxiliary manner until the flatness of the mask coiled material in the scanning area is qualified. According to the technical scheme provided by the invention, the flatness can be quickly improved in the mask coiled material conveying process, and the mask coiled material production efficiency is improved.

The application discloses a millimeter wave radar distance domain target detection method based on complex binary transformation, and realizes target distance detection by constructing a double-path symmetric complex binary network, and the implementation scheme comprises the following steps: 1. constructing samples containing different target distances and noise intensities as a training data set; 2. constructing a complex binary transformation network (CVBT) model; 3. constructing a loss function by adopting a GrandNorm strategy, solving the gradient disappearance problem in the back propagation of the binary neural network by using a straight-through estimator, and training the CVBT network parameters by using the training sample data set to obtain the CVBT network parameters; 4. performing complex binary transformation on radar echo by using the trained CVBT network, obtaining a distance spectrum, and performing constant false alarm rate detection on the distance spectrum to obtain the specific position of the target. The method realizes the purpose of replacing FFT for frequency domain analysis by adopting the binary neural network parameter training mode for the first time.

The application provides a system for adjusting flatness of a mask roll. The system comprises a horizontal straight transmission area of the mask roll formed by a plurality of conveying rollers, a flatness adjusting roller group, a line laser scanner and a computing processing module. All rollers of the flatness adjusting roller group press against the segmented area of the mask roll to completely cover the width of the mask roll. The computing processing module continuously obtains a reflected light band deformation image of the line laser on the upper surface of the mask roll in the scanning area through the line laser scanner and inputs the image into a pre-trained lightweight binary neural network to determine whether the flatness of the mask roll in the scanning area is qualified. When the flatness of the mask roll is unqualified, the height of each roller of the flatness adjusting roller group is adjusted according to the continuous cross-sectional profile of the upper surface of the mask roll in the scanning area until the flatness of the mask roll in the scanning area is qualified. The technical scheme provided by the application can quickly adjust the flatness of the mask roll during the transmission process and improve the production efficiency of the mask roll.

The invention discloses a data processing method and device of an intermediate neural network layer and a random binary neural network, and relates to the technical field of neural networks, and the method comprises the steps: obtaining the output data of a preorder neural network layer, and converting the output data of the preorder neural network layer into weighted summation data; based on preset sampling times, performing multiple independent Bernoulli sampling on the weighted summation data to obtain multiple sampling results; and based on multiple sampling results, determining target sampling data, and transmitting the target sampling data to a post-order neural network layer for reasoning. Through the above mode, a noise suppression mechanism is moved forward from an output layer to an intermediate layer, and multiple sampling and averaging operations are carried out in the intermediate layer, so that instantaneous noise introduced by single random sampling can be effectively smoothed, the statistical variance output by an activation value of the intermediate layer is reduced, and the stability of signal transmission between layers of a neural network is improved. The cascade amplification effect of noise in the deep network is effectively suppressed, and the overall reasoning precision of the model is improved.

The application provides an optical orbital angular momentum recognition method and system based on a binary neural network model, and aims to solve the technical problem that the OAM recognition method based on a spatial light modulator deep learning method has high calculation complexity, large memory occupation, and is difficult to meet the application requirements of lightweight and real-time of space laser communication. The application provides an optical orbital angular momentum recognition method based on a binary neural network model, which realizes the recognition of different topological charge OAM by constructing a binary neural network model. Compared with the traditional OAM recognition method based on the spatial light modulator deep learning method, the application adopts binary weights and binary activation functions, and the parameter storage requirement is significantly reduced. At the same time, the binary neural network model adopted by the application adopts PopCount processing instead of convolution operation, reduces the calculation amount, enhances the real-time performance, and can meet the requirement of fast recognition in the application scene of inter-satellite communication.