627 results about "Back propagation algorithm" patented technology

The invention discloses a face image enhancement method based on a generative adversarial network (GAN). The method comprises the following steps of: 1, preprocessing face images having multiple posesby using a 3D dense face alignment method; 2, designing a face enhancement network based on the GAN, and generating the GAN in two steps; 3, according to a task requirement, designing an objective function 4 corresponding to the Step-I and the Step-II, pre-training a recognition model by using MS-1-celeb, and pre-training a TS-GAN model by using augmented data; and 5, using Multi-PIE as a training Set, using a back propagation algorithm (4) to complete the pre-trained TS-GAN model parameters until convergence. The finally trained TS-GAN model can obtain a front face image corresponding to aninput image. Further, the front face image retains original illumination, is true in visual degree, and retains original identity information.

The invention relates to an entity relationship joint extraction method and system based on an attention mechanism. The method comprises the following steps of: converting an entity marked in trainingdata and a triple of a relationship into a form that each word corresponds to a predefined type of tag; Mapping each word in the sentences of the training data into a corresponding word vector, inputting the word vectors into a neural network model based on an attention mechanism, and performing training through a back propagation algorithm to obtain a label prediction model; And inputting the sentences needing to be subjected to entity relationship extraction into the trained label prediction model, predicting a label corresponding to each word, and obtaining entity relationship triples existing in the sentences according to the corresponding relationship between the labels and the words in the triples. The system comprises a preprocessing module, a model training module and a result processing module. According to the method, by more effectively utilizing the key information in the sentences, the joint extraction performance of the relational entities is improved, and the method hasgood practicability.

The invention relates to a modeling approach and a modeling system of an acoustic model used in speech recognition. The modeling approach includes the steps of: S1, training an initial model, wherein a modeling unit is a tri-phone state which is clustered by a phoneme decision tree and a state transition probability is provided by the model, S2, obtaining state information of a frame level based on the fact that the initial model aligns the tri-phone state of phonetic features of training data compulsively, S3, pre-training a deep neural network to obtain initial weights of each hidden layer, S4, training the initialized network through error back propagation algorithm based on the obtained frame level state information and updating the weights. According to the modeling approach, a context relevant tri-phone state is used as the modeling unit, the model is established based on the deep neural network, weight of each hidden layer of the network is initialized through restricted Boltzmann algorithm, and the weights can be updated subsequently by means of error back propagation algorithm. Therefore, risk that the network is easy to get into local extremum in pre-training is relieved effectively, and modeling accuracy of the acoustic model is improved greatly.

Systems and methods are provided for fusing new evidence and displaying node parameters of a decision network. The decision network can be a belief network, such as a Dempster-Shafer belief network. The Dempster-Shafer belief network includes node parameters that conform to the Dempster-Shafer combination rule, which is based on an evidential interval: the sum of a belief value, a disbelief value and an unknown value is equal to one. A user override is provided to allow a user to override node parameters associated with a hypothesis or outcome and the decision network self-adjusts the appropriate link values or learned to instantiate the override. The back-propagation algorithm is used to adjust the links.

The invention discloses a mechanical failure migration diagnosis method and system based on adversarial learning. The method comprises the following steps: acquiring and analyzing original signals ofmechanical failure under different working conditions to generate a labeled source domain training dataset, an unlabelled source domain training dataset and a target domain test dataset under different working conditions; training a deep convolutional neutral network model according to the labeled source domain training dataset and a back propagation algorithm to generate a failure diagnosis model; training the failure diagnosis model according to the unlabelled source domain training dataset and the target domain test dataset; fine adjusting the trained failure diagnosis model according to the labeled source domain training dataset and the back propagation algorithm; inputting the unlabelled target domain test dataset into the fine adjusted failure diagnosis model, and outputting the failure category of a to-be-tested sample. By means of the method, the domain invariant feature is obtained with the adversarial learning method, migration among different domains is realized, and intelligent diagnosis of mechanical failure under variable working conditions is realized.

The invention discloses a character detection method and device based on deep learning. The method comprises the steps: designing a multilayer convolution neural network structure, and enabling each character to serve as a class, thereby forming a multi-class classification problem; employing a counter propagation algorithm for the training of a convolution neural network, so as to recognize a single character; minimizing a target function of the network in a supervision manner, and obtaining a character recognition model; finally employing a front-end feature extracting layer for weight initialization, changing the node number of a last full-connection layer into two, enabling a network to become a two-class classification model, and employing character and non-character samples for training the network. Through the above steps, one character detection classifier can complete all operation. During testing, the full-connection layer is converted into a convolution layer. A given input image needs to be scanned through a multi-dimension sliding window, and a character probability graph is obtained. A final character region is obtained through non-maximum-value inhibition.

The invention discloses a dynamic gesture recognition method based on a hybrid deep learning model. The dynamic gesture recognition method includes a training phase and a test phase. The training phase includes first, training a CNN based on an image set constituting a gesture video and then extracting spatial features of each frame of the dynamic gesture video sequence frame by frame using the trained CNN; for each gesture video sequence to be recognized, organizing the frame-level features learned by the CNN into a matrix in chronological order; inputting the matrix to an MVRBM to learn gesture action spatiotemporal features that fuse spatiotemporal attributes; and introducing a discriminative NN; and taking the MVRBM as a pre-training process of NN model parameters and network weights and bias that are learned by the MVRBM as initial values of the weights and bias of the NN, and fine-tuning the weights and bias of the NN by a back propagation algorithm. The test phase includes extracting and splicing features of each frame of the dynamic gesture video sequence frame by frame based on CNN, and inputting the features into the trained NN for gesture recognition. The effective spatiotemporal representation of the 3D dynamic gesture video sequence is realized by adopting the technical scheme of the invention.

The invention discloses a method and a device for data identification based on a multitask deep neural network. The method comprises the following steps of: step 1, establishing the multitask deep neural network; step 2, regarding the two adjacent layers of the multitask deep neural network as a limited Boltzman machine, and pre-training the multitask deep neural network by a method of training layer by layer without supervision from bottom to top, so as to obtain an initial connection weight between the adjacent layers; step 3, minimizing a target function about the network weight with supervision by virtue of a back propagation algorithm, so as to obtain an optimized network weight; and step 4, inputting to-be-identified data in the multitask deep neural network with the optimized network weight, so as to obtain an output layer node value, thus obtaining the type of the to-be-identified data according to the output layer node value. In the method, relevance among different labels is excavated by virtue of the neural network, so that a high image labelling accuracy can still be ensured in large-scale image labelling with a high label quantity.

The invention discloses a real-time high-performance street-view image semantic segmentation method based on deep learning. The real-time high-performance street-view image semantic segmentation method includes the steps: preparing a street-view image training, verifying and testing data set; carrying out downsampling on images of the data set to reduce the resolution of the images; transforming an existing lightweight classification network to serve as a basic feature extraction network of semantic segmentation; connecting identification hole space pyramid pooling in series after the basic feature extraction network for solving the multi-scale problem of semantic segmentation; stacking a plurality of convolutional layers to form a shallow spatial information storage network; fusing the obtained feature maps by using a feature fusion network to form a prediction result; comparing the output image with a semantic annotation image in the data set, and performing end-to-end training by using a back propagation algorithm to obtain a real-time high-performance street-view image semantic segmentation network model; and inputting the street-view image to be tested into the real-time high-performance street-view image semantic segmentation network model to obtain a semantic segmentation result of the street-view image.

The present invention discloses a facial emotion recognition method based on a depth sparse self-encoding network. The method comprises the steps of 1, acquiring and pre-processing data; 2, establishing a depth sparse self-encoding network; 3, automatically encoding / decoding the depth sparse self-encoding network; 4, training a Softmax classifier; and 5, finely adjusting the overall weight of the network. According to the technical scheme of the invention, sparseness parameters are introduced. In this way, the number of neuronal nodes is reduced, and the compressed representation of data can be learned. Meanwhile, the training and recognizing speed is improved effectively. Moreover, the weight of the network is finely adjusted based on the back-propagation algorithm and the gradient descent method, so that the global optimization is realized. The local extremum and gradient diffusion problem during the training process can be overcome, so that the recognition performance is improved.

The invention discloses a convolution neural network based face detection method and apparatus. The method comprises the steps of: zooming all pictures in a training set into grayscale images of a predetermined size, and assigning label information to each pixel point of each picture, wherein the label information is used to show whether the pixel point that corresponds to the information is a face; establishing a convolution neural network, wherein layers of the convolution neural network are sequentially an input layer, a plurality of convolution layers, a plurality of full connectivity layers and an output layer; training the convolution neural network by using a gradient descent method and a back propagation algorithm; inputting a photograph to be detected into the well trained convolution neural network, so as to obtain an output characteristic value of a final layer; comparing the output characteristic value of a final layer to a predetermined threshold, so as to determine whether each pixel point of the photograph is a face region; and using a method of minimum closure, and detecting a face position according to pixel points that are determined as the face region.

The invention discloses a speech classification method based on a deep neural network, and aims at solving different speech classification problems through a unified algorithm model. The method of theinvention includes the steps of S1, converting a speech into a corresponding spectrogram; segmenting the complete spectrogram along the frequency domain into blocks to obtain a local frequency domaininformation set; S2, taking the complete and local frequency domain information as inputs of a model respectively, and based on the different inputs, the convolutional neural network being capable ofextracting local and global features; S3, using an attention mechanism to fuse global and local feature expressions to form a final feature expression; S4, using tagged data to train the network by gradient descent and back propagation algorithms; and S5, using the trained parameters for an untagged speech and taking the classification of highest probability that the model outputs as a predictionresult. The method of the invention realizes a unified algorithm model for different speech classification problems, and improves the accuracy on multiple speech classification problems.

The invention discloses an enterprise entity relation extraction method based on a convolutional neural network. The method comprises the steps of a relation corpus building stage, wherein an initial seed relation pair set is built artificially, and by means of an internet search engine and a Bootstrapping technology, relation language materials are generated in an iteration mode, and finally a relation corpus is formed; a relation classification model training stage, wherein term vectors and position embedding are combined to build a sentence vector matrix representation to serve as input of a network, the convolutional neural network is built, the network is trained by means of a back propagation algorithm, and a relation classification model is obtained; an enterprise entity relation extraction stage in a web page, wherein the web page is preprocessed by combining web page text extraction with a named entity identification technology, and then enterprise entity relation extraction is conducted on the preprocessed web page. By means of the method, not only the defects of an artificial feature method can be overcome, but also the enterprise entity relation can be extracted from the web page more accurately and efficiently.

The invention provides a fine-grained human face image fast retrieval method based on deep learning. The method comprises the following steps that a deep convolution neural network model is established, and a loss layer used for calculating a loss function is added to each branch layer of the deep convolution neural network model; parameters of the deep convolution neural network model are initialized; a data set is constructed, and according to a preset proportion, pictures in the data set are randomly divided into a training set, a test set and a verification set; learning parameters of the depth convolution neural network model are set; the deep convolution neural network model is trained, and the parameters of the deep convolution neural network model are updated through a stochastic gradient descent and back propagation algorithm; the trained deep convolution neural network model is tested and is specifically subjected to a coarse-grained test and a fine-grained test, human face retrieval results are obtained, and then the image retrieval efficiency and the image retrieval precision are improved.

The invention discloses a gastrointestinal tumor microscopic hyper-spectral image processing method based on a convolutional neural network, comprising the following steps: reducing and de-noising the spectral dimension of an acquired gastrointestinal tissue hyper-spectral training image; constructing a convolutional neural network structure; and inputting obtained hyper-spectral data principal components (namely, a plurality of 2D gray images, which are equivalent to a plurality of feature maps of an input layer) as input images into the constructed convolutional neural network structure using a batch processing method, and by taking a cross entropy function as a loss function and using an error back propagation algorithm, training the parameters in the convolutional neural network and the parameters of a logistic regression layer according to the average loss function in a training batch until the network converges. According to the invention, the dimension of a hyper-spectral image is reduced using a principal component analysis method, enough spectral information and spatial texture information are retained, the complexity of the algorithm is reduced greatly, and the efficiency of the algorithm is improved.

The invention belongs to the technical field of neural network, and provides a method and apparatus for increasing generalization capability of a convolutional neural network. The method includes the following steps: reading a group of images from a training set, mapping the group of images to a plurality of image character vectors, dividing the plurality of image characteristic vectors into a plurality of classes based on the types of the images; based on the image characteristic vectors of each class, calculating the intra-class loss function of all the characteristic vectors; based on the image characteristic vectors of each class, calculating the inter-class loss function of all the characteristic vectors; based on the intra-class loss function of all the characteristic vectors, using the back propagation algorithm to update the weight of each node of the convolutional neural network; repeating the above mentioned steps until the convolutional neural network converges on the training set or reaches preset repeating times. According to the invention, the method and the apparatus can save all data in long-tailed distribution, makes full usage of abundant inter-class information of tail data, and increases the generalization capability of the convolutional neural network.

The invention discloses a neonatal pain expression classification method based on a CNN (convolutional neural network). The method comprises the following steps: firstly, neonatal pain expression images are acquired, the images are classified by professional medical staff according to calm, cry, slight pain and intense pain level by level, and a neonatal pain expression image library is established; secondly, the CNN containing one data layer, three convolutional layers, two full connection layers and one classification layer is constructed; then samples in the neonatal pain expression image library are taken as data input of the CNN, iterative training is performed on the CNN with a back propagation algorithm, and optimizing training is performed on global parameters, so that a network output loss function value is reduced and converged; finally, neonatal pain expression test samples are input and then recognized and classified with the CNN, expression recognition of neonatal infants in calm, cry, slight pain and intense pain states is realized, and one novel method and way is provided for evaluating neonatal pain degree.

The invention discloses a single image super-resolution reconstruction method based on a depth component learning network, which comprises the following steps: expanding a training sample image and carrying out region extraction and degradation operations to obtain corresponding high-resolution and low-resolution image training sets; a deep network with component learning structure being constructed. The network decomposes the input low-resolution image into global components, and then predicts the corresponding image in high-resolution space using the residual components extracted from the global components. Batch random gradient descent method and back propagation algorithm are used to train the constructed deep component network iteratively on the training set, and the model with optimized weights is obtained. Reconstruction of Low Resolution Images Using the Trained Component Networks; the reconstruction result is restored to the original color space, and the final output of the super-resolution reconstruction is obtained. The method of the invention not only can improve the quality of the reconstructed super-resolution image, but also can improve the operation speed of the model.

The invention provides a method for identifying an abnormal sound signal based on a convolutional neural network, and relates to the technical field of sound signal classification and identification.The method comprises the steps that firstly, six abnormal sound samples are collected by using an existing abnormal sound bank to form a sample sound bank and form samples with noise; then the sound in the sample sound bank is preprocessed and arranged in two dimensions of time and frequency domain into a two-dimensional sound feature graph as the input of a convolutional neural network model; theerror between an actual output result of a training set and a label result is calculated by using a cost function, a difference value is transferred by using a back propagation algorithm, and a weight vector in a full connection layer of the convolutional neural network is updated; the convolutional neural network model is trained by using a supervised learning method; lastly, data in a test setis input, and the accuracy of the convolutional neural network model is verified. The method for identifying the abnormal sound signal based on the convolutional neural network can identify the abnormal sound signal more efficiently and accurately.

The invention provides a character recognition method based on deep learning. The character recognition method based on deep learning comprises the steps that a deeper multilayer convolution neural network structure is designed, and each character acts as a class; the convolution neural network is trained by using a back propagation algorithm so as to recognize the single character, the target function of the network is minimized in a supervised way and a character recognition model is obtained; and finally the most possible words are found out of the dictionary by adopting a viterbi algorithm according to the existing recognized characters. One input is given, candidate characters are obtained by performing sliding window scanning firstly and then the most possible words are found out of the candidate characters in testing. The character features are learned by utilizing the deeper convolution neural network, the method has robustness in character color, size, illumination and fuzziness, and relatively high accuracy of character recognition and word recognition can be maintained.

The invention provides a photovoltaic module fault diagnosis method of a deep convolution generative adversarial network. The method comprises the steps of establishing a mathematical model of a photovoltaic module; carrying out fault image acquisition on the photovoltaic module; setting a part of fault data as a training sample; constructing a training model of the deep convolutional adversarialnetwork; the generator G inputting a noise vector and outputting a pseudo image through a deconvolution layer; the discriminator D inputting a real sample and a pseudo sample, extracting convolution features through convolution operation, and obtaining the probability of the real sample; optimizing a weight parameter through a back propagation algorithm, then starting the next cycle, and outputting a test image every 300 cycles; and inputting the real sample and the obtained test sample into a classifier to classify fault types, thereby realizing fault diagnosis. According to the fault diagnosis method, a large number of fault pictures are generated by using the deep convolutional network, and a fault image database is expanded, so that fault classification is more detailed, and fault diagnosis is more accurate.

The invention discloses an image content question and answer method based on multi-modality low-rank dual-linear pooling. The image content question and answer method comprises the following steps of 1, conducting data preprocessing on an answer text of a question described with an image and a natural language; 2, conducting characteristic fusing on a multi-modality low-rank dual-linear pooling model; 3, establishing a neural network structure based on an MFB pooling model and a collaborative concern model; 4, training the models and utilizing a backward propagation algorithm to train neural network parameters. The neural network model aiming at an image question & answer is provided, and especially a method for conducting unified modeling on cross-media data in a question-answer in the field of the image question and answer and a network structure for studying the 'collaborative concern' to conduct modeling description on an image and question fine-grained characteristics are provided, and the best effect currently in the field of the image question and answer is obtained.

The embodiment of the invention discloses an image labeling method, system, and device and a computer-readable storage medium. The method comprises: a training sample set and a to-be-labeled image into a convolutional neural network, wherein the training sample set employs a back propagation algorithm to minimize a cross entropy loss function in the convolutional neural network to adjust the weight of the convolutional neural network for training and the weight of the trained convolutional neural network is loaded again to extract a sample network feature set of the training sample set and a test network feature set of the to-be-labeled image; according to the sample network feature set, the test network feature set, and a label set, the probability of belong to each kind of label in the to-be-labeled image is calculated and thus a label probability set is generated; and according to the label probability set, labeling is carried out on the to-be-labeled image. On the basis of combination of the deep learning and label propagation algorithms, the high-level semantic features of images are extracted automatically, so that the efficiency and accuracy of image labeling are improved.

The invention discloses an automatic license plate identification method based on a deep convolutional neural network. The automatic license plate identification method comprises the steps of: firstly, designing a network structure and an input format of the neural network; adopting random affine transformation to synthesize a training sample, synthesizing a real scene picture and a grey-scale image license plate, adding noise to simulate and generate a large number of license plate images in a real scene; subjecting the neural network to back-propagation training, and training the neural network by adopting a supervised back-propagation algorithm; conducting sliding window searching, positioning a license plate through sliding a window, segmenting a picture and converting the picture into grey-scale images, and standardizing the grey-scale images to the standard input format. The automatic license plate identification method can effectively handle the influence on identification caused by image translation and rotation, can avoid the dependence on the specific environment and font in the identification process, is simple in algorithm implementation and high in robustness, and is easy to transplant.

The invention relates to a human-shaped image segmentation method. The method includes: for all first pixel points for training a human-shaped image, extracting multi-scale context information; putting image of all scales of all the first pixel points into the same convolutional neutral network, and forming a multichannel convolutional neural network group, each channel corresponding to an image block of a scale; adopting a back propagation algorithm to train the neural network group, and obtaining human-shaped image segmentation training model data; for all second pixel points for training the human-shaped image, extracting multi-scale context information; putting image block of different scales of each second pixel point into a neural network channel corresponding to the human-shaped image segmentation training model, if a first probability is larger than a second probability, the second pixel pints belonging to a human-shaped region, and otherwise, the second pixel points belongs to outside the human-shaped region. The human-shaped image segmentation method in the invention is fast in image segmentation speed and high in accuracy.

The invention provides a monocular depth estimation method based on deep learning, and the method is based on an unsupervised convolutional neural network structure for monocular depth estimation, andcomprises an encoder, a multi-scale feature fusion module, a gating adaptive decoder, and a refinement unit. The method comprises the following steps: S1, preprocessing a data set; S2, constructing aloss function of the convolutional neural network, inputting a training set image, calculating a loss value of the loss function by using a back propagation algorithm, and performing parameter learning by reducing errors through repeated iteration to enable a prediction value to be close to a real value so as to obtain an optimal weight model of the convolutional neural network; and S3, loading the weight model trained in the step S2, and inputting the test set into an unsupervised convolutional neural network for monocular depth estimation to obtain a depth prediction image. The method solves the problems of large calculation amount during offline training and poor detail part recovery effect in deep reconstruction.

The invention, which belongs to the technical field of image processing and computer vision, relates to a sitting posture detecting method based on target detection and body posture estimation. A fusion feature formed by fusion of a feature I and a feature II is extracted, the fusion feature is inputted into a CNN, and if the fusion feature is from a training set, the feature is used for traininga network parameter; if the fusion feature is from a verification set, the feature is used for verifying a network parameter, an error signal is transmitted by a back propagation algorithm, the gradient is updated, an optimal value is found, and classification regression is carried out by using flexible maximum activation function Softmax to obtain a final classification result and classificationaccuracy. Therefore, a problem of target loss under the complicated condition with multiple targets in existing sitting posture detection is solved; the traditional method relying on a wearable deviceor sensor is abandoned; and with the method based on target detection and body posture estimation, the sitting posture of each task target can be determined accurately under the conditions of complicated background and high group density.

A triaxial magnetic electronic compass error compensation method based on depth learning. An implicit error model is trained for compensating non-linear error in magnetic compass measurement and improving the orientation accuracy of the magnetic compass. The error model training consists of two stages: a first stage is pre-training, and a second stage is reverse trimming by using a back-propagation algorithm for fine-tuning all layers of the network, and reducing the error of model training. The magnetic compass calibration and compensation procedure is to use depth learning algorithm training to obtain a non-linear error model, and the distorted measurement magnetic field is inversed back to a true magnetic field value, thereby reducing the calculation error of course angle. The invention aims at nonlinear error of magnetic compass and provides the error training method based on depth learning; compared to random initialization of a traditional neural network, the weight of each layer locates in a better position of parameter space, so as help to improve the convergence of the algorithm and model training accuracy and achieve high-precision orientation of magnetic compass.

The present invention relates to a pipeline abnormality type detection method based on deep learning. The method comprises: extracting video frames from the video captured by a pipeline robot to forma pipeline image set and an image tag set; dividing the pipeline abnormality types into 7 situations of normal pipelines, pipeline staggered joints, pipeline ring cracks, pipeline long cracks, tree roots entering, accumulated floats, and water in the pipeline; establishing a training data set, dividing pictures into a training set, a verification and test set, as well as a corresponding tag set; and using the produced data set to train the convolutional neural network through the error back-propagation algorithm BP, and outputting the probability that the image corresponds to 7 abnormality types.

The invention belongs to the technical field of condition monitoring and fault diagnosis of a wind turbine generator system, in particular to a migration diagnosis method of the gearbox fault of a wind turbine generator system. The method comprises the following steps: establishing four neural network structures, namely, a source domain feature extractor, a target domain feature extractor, a domain classifier and a domain discriminator; obtaining predictive label values by forward propagation from annotated source domain data, network training loss functions are calculated according to predictive label and actual label, and source domain feature extractor and domain classifier are pre-trained by back propagation algorithm. The loss functions of the source domain feature, the target domainfeature and the domain discriminator are calculated by forward propagation from the source domain data and the target domain data, and the domain discriminator and the target domain feature extractorare trained by back propagation algorithm respectively. The newly acquired target domain data is input into the target domain feature extractor, the feature is calculated, and the predictive label ofthe new data is obtained by the domain classifier input from the feature.