859 results about "Non linear mapping" patented technology

The invention discloses a comment text emotion classification model training and emotion classification method and device and equipment and belongs to the field of text emotion classification in natural language processing. Model training comprises the steps that a comment text and associated subject and object information are acquired; a comment subject and object attention mechanism is fused based on a first-layer Bi-LSTM network to extract sentence-level feature representation; the comment subject and object attention mechanism is fused based on a second-layer Bi-LSTM network to extract document-level feature representation; and a hyperbolic tangent non-linear mapping function is adopted to map document-level features to an emotion category space, softmax classification is adopted to train parameters in a model, and an optimal text emotion classification model is obtained. According to the method, the hierarchical bidirectional Bi-LSTM network model and the attention mechanism are adopted, context semantic robust perception and semantic expression of the text can be realized, the robustness of text emotion classification can be remarkably improved, and the correct rate of classification is increased.

The invention belongs to the field of image processing, aims at restoring a high-resolution depth image and utilizing the great learning capacity of convolutional neural networks to solve the defects that the conventional algorithm is high in computational complexity and high in actual application cost and cannot effectively extract features, and provides the technical scheme of a depth map super-resolution reconstruction method based on the convolutional neural networks. The convolutional neural networks (CNN) combining a convolutional layer and a deconvolutional layer is utilized to extract the depth image features of low-resolution sample depth image block and a high-resolution sample depth image block, and then the nonlinear mapping relation between the depth image features is learnt so as to restore the high-resolution depth image. The depth map super-resolution reconstruction method based on the convolutional neural networks is mainly applied to the occasion of image processing.

Methods and apparatuses for nonlinear scaling of video images. To match the aspect ratios of a video image and the target display area, at least one embodiment of the present invention scales the video image according to one or more nonlinear functions along the horizontal direction and / or the vertical direction. In one embodiment, the nonlinear functions are such that the original aspect ratio of the video image is preserved near the center region (or strip) of the image and the image is gradually stretched (or compressed) as it is mapped to the edges. In one example, the scaling is implemented by the texture mapping functionality of OpenGL using graphics hardware. In one embodiment of the present invention, the nonlinear mapping is constructed according to a polynomial mapping; and, the coefficients of the polynomial are adjustable by a user to trade off distortion between the image center and the image edges, giving the user control over the location and the amount of distortion.

A method, system, and computer product is presented for mapping a set of patterns into an m-dimensional space so as to preserve relationships that may exist between these patterns. A subset of the input patterns is chosen and mapped into the m-dimensional space using an iterative nonlinear mapping process based on subset refinements. A set of n attributes are determined for each pattern, and one or more neural networks or other supervised machine learning techniques are then trained in accordance with the mapping produced by the iterative process. Additional input patterns not in the subset are mapped into the m-dimensional space by determining their n input attributes and using the neural networks in a feed-forward (prediction) mode.

The invention discloses a building air-conditioning energy consumption prediction method based on a BP neural network model. The method comprises: analyzing influence factors of building air-conditioning energy consumption; according to influence parameters, collecting historical building air-conditioning energy consumption sample parameters, and preprocessing the parameters; using a BP neural network, according to dimensionality of the sample parameters, establishing a building air-conditioning energy consumption prediction model; using the preprocessed sample parameters as a training sample, training the building air-conditioning energy consumption prediction model; collecting near-term real-time building air-conditioning energy consumption sample parameters to evaluate the building air-conditioning energy consumption prediction model; if errors are in an allowed range, output of the model being a building air-conditioning energy consumption predicted value; if not, training the model again. The building air-conditioning energy consumption prediction method based on a BP neural network model is advantaged in that learning rules are simple, a computer can easily implement, and the method has excellent robustness, memory capability, nonlinear mapping capability, and powerful self-learning capability.

The invention discloses a memristor-based neuron circuit, comprising a synaptic circuit, a neuron activation function circuit and a synaptic weight control circuit. In the synaptic circuit, a memristor, under the control of four MOS tubes, changes the memristor value to simulate the change of the synaptic weight in the neuron network. The designed neuron synaptic circuit is capable of being directly connected with a digital logic electrical level so as to achieve convenient and real-time adjustment to the synaptic weight and through the use of the feature that the output voltage of an operational amplifier is restricted by the power supply voltage, the neuron circuit activation function can be realized as a saturated linear function. The neuron synaptic weight change circuit can utilize the existing CMOS micro-controller and at the same time, the micro-controller can be loaded by the neuron network algorithm to change the synaptic weight to realize corresponding functions. According to the invention, a plurality of neuron circuits could be connected into a large-scale neuron network for complicated functions such as mode identification, signal processing, associated memory and non-linear mapping, etc.

The invention provides a data processing method of a neural network processor and a neural network processor. The method includes the following steps that: input data and corresponding weight absolute values are added together through an adder, wherein the input data are data of the output of a previous stage, and the input data and the weight absolute values are n-element vectors; and n-term data obtained after adding the input data and the corresponding weight absolute values together are subjected to n times of first nonlinear mapping; results obtained after the first nonlinear mapping are subjected to n times of accumulation operation through an accumulator, wherein the accumulation operation includes weight sign bit-controlled adding operation and subtraction operation; and a result obtained after the n times of accumulation operation is subjected to second nonlinear mapping, so that a processing result can be obtained, and data output is carried out, wherein the second nonlinear mapping is formulated according to the rules of neural network nonlinear mapping and the inverse mapping of the first nonlinear mapping. With the method adopted, quantization efficiency can be improved, and storage requirements and bandwidth requirements of data can be decreased.

A radio frequency amplifier system is disclosed in which the amplifier bias supply and power supply voltages are >instantaneously modulated with signals derived from the envelope voltage of the input signal. Separate non-linear mapping functions are used to derive the supply and bias voltages. The two mapping functions are optimised jointly to achieve particular system performance goals, such as optimum efficiency, constant gain, constant phase, or minimum spectral spreading. An optimisation of >particular interest is that which achieves best RF amplifier power added efficiency subject to achieving constant amplifier gain. In this way the need for pre-distortion linearization may be reduced or eliminated.

The invention provides a magnetic resonance image feature extraction and classification method based on deep learning, comprising: S1, taking a magnetic resonance image, and performing pretreatment operation and feature mapping operation on the magnetic resonance image; S2, constructing a multilayer convolutional neural network including an input layer, a plurality of convolutional layers, at least one pooling layer / lower sampling layer and a fully connected layer, wherein the convolutional layers and the pooling layer / lower sampling layer are successively alternatively arranged between the input layer and the fully connected layer, and the convolutional layers are one more than the pooling layer / lower sampling layer; S3, employing the multilayer convolutional neural network constructed in Step 2 to extract features of the magnetic resonance image; and S4, inputting feature vectors outputted in Step 3 into a Softmax classifier, and determining the disease attribute of the magnetic resonance image. The magnetic resonance image feature extraction and classification method can automatically obtain highly distinguishable features / feature combinations based on the nonlinear mapping of the multilayer convolutional neural network, and continuously optimize a network structure to obtain better classification effects.

The invention discloses an image enhancement method on the basis of an improved multi-scale Retinex theory. The method comprises the following steps of: carrying out nonlinear adjustment on the details of dark areas and the brightness of highlighted areas by virtue of a global brightness adjustment function; enhancing an image by virtue of a canonical gain compensation multi-scale Retinex algorithm; and according to the mean brightness value of a selected area, calculating the parameters of an S curve, adaptively adjusting the S curve, and carrying out the procedures of nonlinear mapping and the like on the enhanced image, thus finishing the enhancement on the image. The method disclosed by the invention solves the problems that when the conventional multi-scale Retinex theory method is used, a halo phenomenon is caused, the overall brightness of an enhanced high dynamic range image is insufficient, and the contrast ratio of the image is low. According to the invention, the S curve is self-adaptively adjusted according to the brightness of the central area of the image, and then the nonlinear mapping is performed on the image, so that the gradation of the image is stretched, and the contrast ratio of the image is improved; and the robustness of the algorithm on a complex night vision image is improved.

The invention discloses a single-frame image super-resolution reconstruction method on the basis of deep learning. The single-frame image super-resolution reconstruction method comprises the following steps: 1, firstly, acquiring characteristics of low-resolution and corresponding high-resolution image blocks by training two automatic encoders; 2, on the basis of the acquired characteristics of the high-resolution and low-resolution image blocks, then training a single-layer neural network and learning a nonlinear mapping relation of two characteristics; 3, on the basis of two automatic encoders and the single-layer neural network, constructing a three-layer deep network, using the low-resolution image block as an input, using the high-resolution image block as an output and finely regulating parameters of the three-layer deep network; 4, according to the obtained three-layer deep network, carrying out single-frame image super-resolution reconstruction, and obtaining the output, i.e. a gray value corresponding to the high-resolution image block, by using a gray value of the low-resolution image block as the input. According to the single-frame image super-resolution reconstruction method on the basis of deep learning, not only is quality of a super-resolution reconstructed image improved, but also super-resolution reconstruction time is shortened and the real-time requirement can be met basically.

The invention discloses an intelligent combination forecasting method for short-term traffic flow, which at least comprises the following modules: 1) a historic average module, which divides different dates of one year into three different types, namely a holiday type, a weekend type as well as a Monday and Friday type, and calculates different traffic flows for each type respectively based on statistical data of historic traffic flow; 2) a neural network module, which is a parallel distributed information processing network, has the functions of non-linear mapping and associative memory, and forecasts the traffic flows through the internal relation modeling of the data; and 3) a fuzzy combination module, which aims at different traffic conditions to perform fuzzy combination transformation on the output of the historic average module and the neural network module to forecast the short-term traffic flow. The forecasting precision of the method is higher than the precision that single forecasting methods are independently used respectively, and is an effective method for forecasting the short-term traffic flow.

An image processing system for implementing a variant of Retinex algorithm based on local illumination correction is provided to improve an intensity value of each position of an input image. The image processing system provides the ability to distinguish details when compressing dynamic range of a given image. Specifically, first the illumination of the image is estimated by performing a weighted low pass filter (WLPF) operation, and the original image is separated into the illumination and the reflectance images. Next, the estimated illumination is modified by using a non-linear mapping function based on local illumination at each pixel location. Then, the modified illumination image is combined with the reflectance image. Finally, an enhanced image is obtained after a contrast correction procedure.

The invention discloses an FOG based on neural network to do temperature compensation, using the neural network learning algorithm for the temperature drift compensation method, and the temperature compensation method obtains the amendment temperature compensation coefficient to provide parameters for FOG high precision output. According to the effect of the temperature to the FOG output when the FOG working, it respectively installs a temperature sensor A on the light source, a temperature sensor B at the Y waveguide, a temperature sensor C at the fiber ring inside and a temperature sensor D at the fiber ring outside, and through temperature change of the four temperature sensors, and using the neural network learning algorithm for temperature compensation model training, it makes the built temperature drift compensation having good nonlinear mapping ability, self-learning ability and the generalization ability; the temperature drift compensation coefficient measurement is measured in the debugging phase, reducing the gyro computing volume when using.

The invention discloses a software failure positioning method based on machine learning algorithm to solve the technical problem of low positioning efficiency of existing software failure positioning methods. According to the technical scheme, the method comprises the steps of describing failure distribution possibly existing in an actual program based on Gaussian mixture distribution to enable failure distribution in the program to be more definite; removing redundant test samples with a cluster analysis method based on a Gaussian mixture model, and finding a special test set for a specific failure, so that the adverse effect of redundant use cases on positioning precision is reduced; remodifying a support vector machine model to be adapted to an unbalanced data sample, and finding the nonlinear mapping relation between use case coverage information and an execution result by means of the parallel debugging theory, so that machine learning algorithm is free from the local optimal solution problem caused by uneven samples; finally, designing a virtual test suite, placing the virtual test suite in a well trained model for prediction, obtaining a statement equivocation value ranking result, and conducting failure positioning. In this way, software failure positioning efficiency is improved.

The invention discloses a rockburst grade predicting method based on an information vector machine to mainly solve the problem that in the current underground engineering construction process, the rockburst geological disaster prediction effect is not good. A rockburst evaluation index and a grading standard are selected, and domestic and overseas great deep rock engineering rockburst instances are widely collected to establish an abundant training sample library. A cross validation strategy is utilized for training an IVM model with the superior statistics mode recognition performance, accordingly, a nonlinear mapping relation between the rockburst evaluation index and the rockburst grade is established, model initial parameter setting and the training sample library are adjusted according to the training result, and the IVM model predicting the rockburst grade is finally established. By means of the method to predict the rockburst grade, the complex mechanical analysis or calculation is not needed, only the input feature vector of a sample to be predicted needs to be input in the prediction model, and then the prediction value of the rockburst grade can be obtained. The method is economical, efficient and high in prediction precision and has good engineering application prospects.

The present invention discloses a depth image denoising and enhancing method based on deep learning. The method comprises the steps of establishing a depth image denoising and enhancing convolutional neural network, wherein the network is composed of three layers of convolution units which finish the functions of feature extraction, non-linear mapping and image reconstruction of the input images respectively; jointly using the depth and visual images as the input of the convolutional neural network, wherein firstly the visual images are processed into the grayscale images in a grayscale processing manner; and enhancing the edge information and taking out the redundant information by the image preprocessing; segmenting the depth images into the image blocks according to certain intervals, adding the effective data by a rotation and pixel overturning data amplification method, and discarding the interference blocks and the redundant blocks; and improving the learning efficiency of the network adaptively based on a loss training depth image enhancement convolutional neural network of a weight map. According to the method of the present invention, the black spot filling and the denoising operations can be carried out on the depth images with noise real-timely, and the good visual effect and depth value recovery effect can be realized.

The invention provides a method and system for optimizing and controlling combustion performance of a circulating fluidized bed boiler in real time. The method for optimizing and controlling combustion performance of the circulating fluidized bed boiler in real time includes the steps that based on historical data of combustion of the boiler, a first neural network used for being matched with a nonlinear mapping relation between the input and the output of an existing fluidized bed boiler system is established; according to the first neural network, the adjustable input variable of the input of the fluidized bed boiler system is determined; and the adjustable input variable serves as the bottom layer control input, the trends of corresponding deviation and the deviation change rate are calculated, and a second neural network is established for training and acquiring the process variable control output. By the adoption of the method and system, the boiler efficiency is improved, the power supply coal consumption is lowered, coking and slag bonding are prevented or treated, NOx emission is reduced, and the safety, the reliability and the economical efficiency of operation of the boiler are further improved.

The invention belongs to the technical field of image processing and provides a neural network-based image adaptive enhancement method. The method uses a neural network to establish a model of nonlinear mapping between an average value and a standard deviation of an image and the enhancement factor of an original image and a high-frequency component of the image. The method comprises the following steps for image adaptive enhancement: calculating the average value and the standard deviation of the image and obtaining the enhancement factor by establishing the nonlinear mapping model; filtering the average value of the image and obtaining a low-frequency component of the image; obtaining the high-frequency component of the image through a difference value of the original image and the low-frequency component; and superposing the high-frequency component and the original image which are multiplied with the enhancement factors respectively to realize the adaptive enhancement of the image. The image adaptive enhancement method has the advantages of achieving small calculation amount and strong real-time, automatically acquiring the enhancement factor according to the average value and the standard deviation of the image, realizing the adaptive enhancement of the brightness and the contrast of the image, remarkably improving the visual effect of low-contrast and low-brightness images, and laying a foundation for image identification.

The invention relates to an image super-resolution reconstruction method, belongs to the image processing technology field and solves problems that the edge information of an image generated in the prior art is fuzzy, application to multiple magnification times cannot be realized and the reconstruction effect is poor. The method comprises steps that a convolutional neural network for training andlearning is constructed, and the convolutional neural network comprises an LR characteristic extraction layer, a nonlinear mapping layer and an HR reconstruction layer in order from top to bottom; inputted paired LR images and HR images are trained through utilizing the convolutional neural network, training of at least two magnification scales is performed simultaneously, and an optimal parameterset of the convolutional neural network and scale adjustment factors at the corresponding magnification scales are acquired; after the training is completed, the target LR images and the target magnification times are inputted to the convolutional neural network, and the target HR images are acquired. The method is advantaged in that the training speed of the convolutional neural network is fast,after training is completed, and the HR images at any magnification times in the training scale can be acquired in real time.

The invention discloses a deep learning-based clustering method. The method comprises the following steps: obtaining the initial network weight of a deep neural network; grouping samples randomly and mapping to a feature space; adding the target function of the original deep neural network into the in-class constraint function of a feature layer; updating the network weight of the deep neural network and calculating to obtain a new feature layer; distributing all the samples to the class group of the nearest clustering center and calculating a new clustering center; substituting the new clustering center for the clustering center of the in-class constraint function, and returning to the network weight updating step to perform iteration to obtain and output the final clustering classification result. According to the method, the samples are subjected to non-linear mapping of the deep neural network from the original data space, which is difficult to cluster, to obtain a height-classified feature for clustering; a better clustering effect can be achieved by continuously optimizing the network structure. The deep learning-based clustering method with lower memory consumption and higher clustering precision is superior to the conventional clustering algorithm.

The present invention relates to a method for analyzing, separating and extracting audio signals. Due to the generation of a series of short-time spectra, a non-linear mapping into the pitch excitation layer, a non-linear mapping into the rhythm excitation layer, extraction of the coherent frequency streams, extraction of the coherent time events and the modeling of the residual signal, the audio signal can be decomposed into rhythm and frequency portions with which the signal can be further processed in a simple manner. The uses of said method are: data compression, manipulation of the time base, tune and formant structure, notation, track separation and identification of audio data.

The invention provides a single image rainfall removing method based on image partitioning of generating antagonistic network. By dividing images into image blocks with the same size and non-overlapping each other, each image block is used as a condition to generate input of antagonistic network, and the input dimension is reduced. A generate antagonism network is trained to realize the nonlinearmapping from rainless image block to rainless image block, which overcomes the problem that many details are neglected and can remove the rainline at every scale as much as possible. In order to maintain the consistency of the structure and color between the rainless image blocks, a new error function is constructed by using bilateral filter and non-mean local denoising algorithm, which is added to the total error function of the conditional generation countermeasure network. The invention does not need any prior knowledge, nor does it need to preprocess and post-process the image, thus ensuring the integrity of the whole structure. The results on the test set show that the invention is improved by 4-7dB compared with the classical algorithm.

A sigma delta modulator includes a modulator module that includes a quantizer with variable hysteresis, which receives an input signal to perform necessary modulation operations. A non-linear mapping module receives a signal associated with said input signal and tabulates the necessary hysteresis control information so as to reduce the transition rate of the modulator module.

The invention discloses a time-dimension video super-resolution method based on deep learning, which is mainly aimed at solving the problems of the prior art that reconstructed video image insert frames are poor in stability and low in precision. The technical key of the method is that fitting of a nonlinear mapping relationship between an original video image and a down-sampling video image is conducted through neural network training. The method comprises the following steps: 1) obtaining an original video image set and a down-sampling video image set and taking the original video image set and the down-sampling video image set as training samples of a neural network; 2) constructing a neural network model, and training parameters of the neural network through the training samples; and 3) taking any given video as a test sample, and inputting the test sample into the trained neural network model, wherein an output result of the neural network is a reconstructed video image. The calculating complexity of reconstruction of video image insert frames is reduced, and the stability and the precision of the reconstructed video image insert frames are improved. The method can be used for scene interpolation and animation making and also for time-domain insert frames of low-frame-rate videos.

The present invention successfully combines neural-net discriminative feature processing with Gaussian-mixture distribution modeling (GMM). By training one or more neural networks to generate subword probability posteriors, then using transformations of these estimates as the base features for a conventionally-trained Gaussian-mixture based system, substantial error rate reductions may be achieved. The present invention effectively has two acoustic models in tandem—first a neural net and then a GMM. By using a variety of combination schemes available for connectionist models, various systems based upon multiple features streams can be constructed with even greater error rate reductions.

The invention discloses a particle swarm optimization neural network model-based method for detecting the moisture content of wood. A particle swarm is combined with a back propagation (BP) algorithm to finish neural network training, so that the training accuracy of a network model is enhanced; and the model is applied to the detection of the moisture content of wood, so that high detection accuracy is achieved. The method has the advantages that: 1) by the properties of randomized global optimization search and high convergence rate of a particle swarm optimization algorithm, overall optimization is performed on the weight of a network, so that the defects of low convergence rate and easy local minimum existing in the BP algorithm are overcome; 2) in the BP algorithm, an approximately optimal weight provided by the particle swarm optimization algorithm is taken as an initial value and further optimization is performed by using the characteristics of nonlinear mapping capability and high local optimization capability of the BP algorithm, so that an optimal value of a network weight is obtained; and 3) the moisture content of wood and an environmental temperature parameter are detected based on an electrical measuring method, a particle swarm optimization neural network model is established and is applied to the detection of the moisture content of wood, and the effectiveness of the method is verified.

The invention discloses a GAN (Generative Adversarial Nets)-based CFA (Color Filer Array) image demosaicing joint denoising method. The method comprises the following steps: (1) a training sample setis acquired; (2) the GAN is built; (3) parameters of a 9-layer convolution neural network are updated; (4) parameters of a 39-layer convolution neural network are updated; (5) whether the updating times of the 39-layer convolution neural network and the 9-layer convolution neural network reach 200 times is judged, if yes, a step (6) is executed, or otherwise, the step (3) is executed; (6) a nonlinear mapping relationship is built; and (7) an image after demosaicing and denoising is acquired. The color information of a noisy CFA image obtained by a digital camera can be well recovered, the noise introduced during an image acquisition process of the digital camera can be effectively suppressed, the appearance of unnatural colors is reduced, and the visual effects of a color image are improved.