1326results about How to "Few parameters" patented technology

The invention provides a multi-scale feature fusion ultrasonic image semantic segmentation method based on adversarial learning, and the method comprises the following steps: building a multi-scale feature fusion semantic segmentation network model, building an adversarial discrimination network model, carrying out the adversarial training and model parameter learning, and carrying out the automatic segmentation of a breast lesion. The method provided by the invention achieves the prediction of a pixel class through the multi-scale features of input images with different resolutions, improvesthe pixel class label prediction accuracy, employs expanding convolution for replacing partial pooling so as to improve the resolution of a segmented image, enables the segmented image generated by asegmentation network guided by an adversarial discrimination network not to be distinguished from a segmentation label, guarantees the good appearance and spatial continuity of the segmented image, and obtains a more precise high-resolution ultrasonic breast lesion segmented image.

The invention relates to a method for relationship classification with an LSTM and CNN joint model based on location. The method includes the steps of (1) preprocessing data; (2) training word vectors; (3) extracting location vectors; acquiring the location vector feature and high-dimensional location feature vector of each word in a training set, cascading the word vector of each word with the high-dimensional location feature vector thereof to obtain a joint feature; (4) building a model for a specific task; encoding contextual information and semantic information of entities by use of bidirectional LSTM; outputting the vector of the location corresponding to the marked entities, inputting the output to CNN, outputting two entity nouns and their contextual information and relational wordinformation, and inputting the entity nouns and their contextual information and relational word information into a classifier for classification; (5) training the model by use of a loss function. The method does not need to manually extract any features, the joint model does not need to use additional natural language processing tools to preprocess the data, the algorithm is simple and clear, and the best effect at present is achieved.

The invention discloses an image classification method based on a semi-supervised self-paced learning cross-task deep network. The method includes the steps of randomly selecting a small amount of labeling samples from the whole image data set, reserving the labels, and remaining all the samples as unlabelled samples having the real labels to be unknown in the whole process, wherein the weight ofthe labeled samples is constant to be one in the training process, the weight of the unlabelled samples is initialized to be zero, and only the labeled samples are used as a training set in the initial process; S2, training a cross-task deep network by the training set; S3, according to the trained cross-task deep network, predicting the pseudo labels of all the unlabelled samples, and giving a corresponding weight of each unlabelled sample; S4, according to a self-paced learning normal form, selecting an unlabelled sample with a high confidence degree, and adding to the training set; and S5,repeating the steps S2-S4 until the cross-task deep network performance is saturated or reaches a preset cycle number. According to the method, the human design feature is not needed to be input, andthe classification can be realized by directly inputting the original image.

The invention discloses a personalized recommendation method based on deep learning. The method comprises the steps of according to the viewing time sequence behavior sequence of the user, predictingthe next movie that the user will watch, including three stages of preprocessing the historical behavior characteristic data of the user watching the movie, modeling a personalized recommendation model, and performing model training and testing by using the user time sequence behavior characteristic sequence; at the historical behavior characteristic data preprocessing stage when the user watchesthe movie, using the implicit feedback of interaction between the user and the movie to sort the interaction data of each user and the movie according to the timestamp, and obtaining a corresponding movie watching time sequence; and then encoding and representing the movie data,wherein the personalized recommendation model modeling comprises the embedded layer design, the one-dimensional convolutional network layer design, a self-attention mechanism, a classification output layer and the loss function design. According to the method, the one-dimensional convolutional neural network technologyand the self-attention mechanism are combined, so that the training efficiency is higher, and the number of parameters is relatively small.

The invention discloses a face identification method and apparatus based on a sequencing neural network model. The method comprises the following steps: performing preprocessing operation on input face images, and correcting angles and expressions of the face images; extracting features of already corrected face images / videos by use of a neural network comprising sequencing operation; and according to feature expression of the face images, calculating similarity between image pairs, and accordingly, obtaining identifies of specific objects in the input face images. According to the invention, a sequencing neural network structure is brought forward for solving the problems of too many parameters and too large calculation cost of a conventional neural network based face identification model in face identification problems, through sequencing expression between different features, network parameters are effectively reduced, and the calculation time is saved; and a training method based on comparison loss and triple loss is brought forward for solving the problem of a quite small number of training data.

The embodiment of the invention discloses a training method and apparatus of a neural network, and an object identification method and apparatus. The training method of a neural network includes the steps: establishing an initial neural network model, wherein the initial neural network model, is an ultra lightweight network SqueezeNet model; acquiring a training sample, and preprocessing the training sample; inputting the training sample which is obtained through preprocessing into the initial neural network model, and according to the first output of determining the loss value of the first output and the preset expectation according to the initial neural network model; and according to the loss value, adjusting the network parameters of the initial neural network model, and generating a target neural network model. The training method of a neural network establishes a high precision target neural network with few parameters and small occupied memory, so as to enable the target neuralnetwork model which is obtained through training to be applied to a mobile terminal, without depending on setting of a GPU (Graphics Processing Unit), thus improving adaptability of the target neuralnetwork model.

The invention belongs to the technical field of computer software and discloses a multi-language-pair neural network machine translation method and system. A plurality of bilingual parallel corpora ofa same language system are utilized and mapped to a same high-dimensional vector space after byte pair encoding, so that multiple languages share a same semantic space, the size of a word list is reduced, model parameters are reduced, and convergence of a model is accelerated. Words of a same language family are in the same vector space, more information can be learned mutually, the information which can not be learned through only certain bilingual parallel corpora can be learnt, and the quality of word vectors is improved. The machine translation system can be used for translation in the language direction without direct bilingual parallel corpora, and the translation quality in the scarce parallel corpus translation direction is greatly improved through mutual information learning. Meanwhile, the same model is used for translation for the translation direction low in utilization rate, occupation of a server is reduced, and the utilization rate of the server is increased.

The invention discloses a multi-view facial expression recognition method based on a mobile terminal, which comprises the steps of cutting out a face region from each image, and carrying out data enhancement to obtain a facial expression recognition method used for training data set of AA-MDNet model; the multi-attitude data set is obtained by GAN model extension, and the multi-attitude data set is obtained by GAN model extension. Using ADN multi-scale clipping method to clip; enter the cropped image into AA-MDNet model, The input image extracts features from DenseNet, a densely connected subnetwork, Then, based on the extracted features, an attention adaptive network (ADN) is used to obtain the position parameters of the attention area of the expression and posture, and the image of the area is scaled from the input image according to the position parameters, which is used as the input of the next scale. Learning the multi-scale high-level feature fusion, we can get the high-level features with global and local fusion features. Finally, we can classify the facial posture and expression categories. The invention has very important significance in the fields of human-computer interaction, face recognition, computer vision and the like.

The utility model discloses a integrated method based on multi-level set evolution and high resolution remote sensing image partition, classification and change inspection, which is characterized in that (1) image preprocessing (radiation, registration and filtering); (2) the multi-level set evolutional partition and classification model, after registration, the GIS data determines the initial profile of each level set function and performs the partition and classification to the first phase image; (3) the model described in the (2) is still adopted, and the initial profile of each level set function is optimized, increment type partition and classification is adopted for the second to T phase; (4) the objective after partition is used as unit, the ith and (i+1)th two adjacent phase image classification results are compared to determine the change area; (5) return back to (3) until the partition, classification and change inspection of all T phase image are finished. The utility model has the advantages that: compared with the traditional pixel-oriented K value method, the classification and inspection precision are improved, The utility model is applicable for the change inspection of sequence remote sensing image and has wide application in hazard monitoring and land resource investigation.

The invention discloses a thrust distribution method for a power positioning system of an offshore drilling platform, which solves the problem of thrust distribution optimization of the power positioning system by applying a particle swarm algorithm through taking minimized propulsion system power consumption and a thrust error as target functions of the thrust distribution optimization and considering the thrust, thrust change rate, azimuth change rate and constriction conditions of a thrust forbidden sector. The method comprises the following steps of setting input quantity of the thrust distribution, including vertical resultant force, transverse resultant force and yawing moment, as given parameters; setting the thrust amplitude and direction of each thruster as unknown variables, namely, solutions to be optimized; analyzing platform stress, wherein the resultant forces and resultant force moments generated by all the thruster are equal to the input quantity of the thrust distribution; and building a thrust distribution mathematical model according to the distribution of offshore drilling platform thrusters and applying the particle swarm algorithm to solve the problem of thrust distribution optimization. The thrust distribution method has no special requirements to the target functions of the thrust distribution and less parameter to be regulated, is simple in operation, easy for realizing and fast in calculation speed, so that the requirement of high real-time capability of the power positioning system is fulfilled.

The invention discloses a power line three-dimensional reconstructing method based on airborne laser radar point cloud. Data of the airborne laser radar point cloud have been correctly classified. The method includes the following steps: 1, loading the correctly-classified data of the airborne laser radar point cloud and initial line track data of an overhead transmission line; 2, extracting accurate information of the positions and the number of electric towers and the track of the power transmission line, and determining the total number of spans of the power transmission line; 3, determining the two-dimensional space range and power line laser radar points of each span of the power transmission line; 4, clustering the determined power line laser radar points of the spans; 5, carrying out power line three-dimensional reconstruction. According to the power line three-dimensional reconstructing method, the number of required parameters is small, the automation degree is high, robustness and universality are better, and the power line three-dimensional reconstructing method has the advantage of being insensitive to the factors such as the number of power lines, the types of the power lines, the power line space arrangement structure, gross error points, point-cloud irregular breaking and the line length; in addition, a reconstruction model has the high reconstruction accuracy.

The invention discloses a group behavior recognition method based on a pseudo 3D convolutional neural network. The group behavior recognition method comprises the following steps: (1) carrying out human body posture estimation and target tracking by utilizing an OpenPose posture estimation algorithm; (2) extracting spatial and temporal features of each person by using a P3D ResNet (pseudo 3D residual network), and classifying the spatial and temporal features by using a softmax classifier to complete single person behavior recognition; (3) constructing a human body target interaction graph byutilizing the position information and the appearance characteristics of the human body target, and performing graph reasoning and training by utilizing a graph convolution network; and (4) performinggroup behavior identification according to the human body target interaction graph. According to the scheme, features are extracted based on a P3D convolutional network; parameters of the model are reduced and identification precision is improved. Through the technology, a computer can judge the state change of the interested human body target in the public area in real time, automatic human bodybehavior recognition is carried out, the recognition precision is high, and the application field is wide.

The invention discloses a high-resolution SAR terrain classification method based on multiscale convolution and feature fusion, and mainly aims at solving the problem in the prior art that the classification precision is low and overfitting easily occurs. The high-resolution SAR terrain classification method comprises the steps of 1, extracting textural features and wavelet features of to-be-classified images; 2, fusing the to-be-classified images, the textural features and the wavelet features to constitute a fusion feature matrix; 3, according to the fusion feature matrix, constructing a training dataset and a testing dataset; 4, adding a multiscale convolution layer and a shuffle layer to an existing CNN network, changing a full-joint layer into a convolution layer, and constructing a multiscale convolution fusion network; 5, using the training dataset to train the multiscale convolution fusion network to obtain model parameters; 6, using the model parameters to initialize the multiscale fusion network to classify a test set. By means of the high-resolution SAR terrain classification method based on the multiscale convolution and the feature fusion, the parameters of the networkare reduced, the overfitting phenomenon of a small sample problem is solved, the classification precision is improved, and the high-resolution SAR terrain classification method can be applied to high-resolution SAR image terrain classification.

The invention discloses a prediction model based hypersonic aircraft sliding-mode control method. The method is used for solving the technical problem of difficulty in engineering realization by existing hypersonic aircraft adaptive discrete control. The method includes: obtaining a strict feedback form of a height subsystem by reasonable assumption, and building a discrete form of an original system through an Eulerian method; building a four-step prediction model of the original system by unceasing forward prediction, wherein the model only includes an equation; enabling the prediction model to provide the relation of height output at a future time and current system state and control input, wherein the relation can be used for computing numerical values of lumped uncertainties of the system at the historical moment, and the numerical values are used for feedback design of a controller; and combining lumped nominal information further, and using a discrete reaching law to design the sliding-mode controller so that robustness of the system is improved. The prediction model based hypersonic aircraft sliding-mode control method has the advantages that the discrete prediction model is built to acquire system nominal and uncertain information, design of virtual control variables is not needed, the controller is simple and practical in design, and the method is suitable for engineering application.

The invention provides an online trajectory planning based efficient bridge crane swing elimination control method. The method includes: constructing a swing angle elimination link for effectively restraining the substantial swing of a load and quickly eliminating a residual swing angle of the load after a trolley reaches a target position; combining the swing angle elimination link with a positioning reference trajectory together in a linear way so as to generate a final trajectory, thus ensuring accurate positioning of the trolley, and at the same time can effectively eliminating the swing angle and the residual swing angle of the load. Compared with existing crane trajectory planning technologies, the method can achieve online generation of a swing elimination trajectory, and has the advantages of simplicity and practicability, small online computation amount, high trolley transportation efficiency, as well as high swing elimination speed, thus being very suitable for real-time control of the crane system.

The invention provides a speech recognition method based on a convolution neural network, which is more good at extracting high-level features, has simple modeling process, is easy to train, has better generalization performance of the model, and can be more widely applied to various speech recognition scenes. The method comprises the following steps: S1, preprocessing the input original speech signal; S2, extracting the key feature parameters reflecting the characteristics of the speech signal to form a feature vector sequence;S 3, base on that DCNN network model, taking the connected time classifier CTC as a loss function, constructing an acoustic model of an end-to-end mode; S4, training the acoustic model to obtain the trained acoustic model; S5, inputting the feature vector sequence to be recognized obtained in the step S2 into the trained acoustic model to obtain a recognition result; and S6, performing a subsequent operation on the basis of the recognition result obtained in step S5, that is, obtaining a word string capable of outputting the speech signal with a maximum probability, that is, a language character after the original speech is recognized.

The invention discloses a tableware shape-based valuation method for a dishes automatic identification system. Dishes are identified through containers as mediums. The method provided by the invention comprises the steps of segmenting the dishes in dinner plates by screening shapes and area characteristics of dishes regions in shot dinner plate images; and then obtaining a classifier by training a convolutional neural network, and directly identifying dishes images to realize dishes identification. Compared with an existing method that is used for only identifying the containers through chips or colors and has limitations of the containers, the tableware shape-based valuation method has the advantages that special tableware does not need to be used as the container, so that the cost is relatively low; and the dishes do not need to be put in the corresponding containers, so that an error introduction factor caused by the fact that the dishes have to be put in the corresponding containers is eliminated.

The invention discloses a brushless direct-current motor stator flux linkage identification method. The method comprises the following steps: the neutral point of a brushless direct-current motor is grounded to make the voltage of the neutral point clamped to zero; the brushless direct-current motor end voltage is determined through the on-off state of a power tube and a fly-wheel diode during the normal conduction process and the commutation process of a three-phase full-bridge inverter, wherein a three-phase phase voltage is obtained by performing subtraction between the three-phase end voltage and the neutral point voltage; detection is performed by adopting a current sensor to obtain a three-phase phase current, and calculation is performed according to the voltage model of the stator flux linkage to obtain a three-phase stator flux linkage; the three-phase stator flux linkage is filtered by a high-pass filter through CLARK transformation; and vector synthesis is performed on the three-phase stator flux linkage filtered by the high-pass filter to obtain a brushless direct-current motor stator flux linkage. According to the invention, advantages of less required motor parameters, simple structure, less calculation amount and low hardware cost can be realized.

The invention discloses a multi-speaker voice separation method based on a convolutional neural network and depth clustering. The method comprises the following steps: 1, a training stage: respectively performing framing, windowing and short-time Fourier transform on single-channel multi-speaker mixed voice and corresponding single-speaker voice; and training mixed voice amplitude frequency spectrum and single-speaker voice amplitude frequency spectrum as an input of a neural network model; 2, a testing stage: taking the mixed voice amplitude frequency spectrum as an input of a threshold expansion convolutional depth clustering model to obtain a high-dimensional embedded vector of each time-frequency unit in the mixed frequency spectrum; using a K-means clustering algorithm to classify thevectors according to a preset number of speakers, obtaining a time-frequency masking matrix of each sound source by means of the time-frequency unit corresponding to each vector, and multiplying thematrixes with the mixed voice amplitude frequency spectrum respectively to obtain a speaker frequency spectrum; and combining a mixed voice phase frequency spectrum according to the speaker frequencyspectrum, and obtaining a plurality of separate voice time domain waveform signals by adopting short-time Fourier inverse transform.

The invention discloses a spectral model transmission method based on unary linear regression. The spectral model transmission method mainly comprises the steps of I. performing the identical preprocessing on sample spectrums of M spectrometers; II. classifying spectrum data into m correction sets and forecasting sets according to m kinds of components of samples, and respectively constructing m correction models on the M spectrometers; III. evaluating the correction models; IV. selecting the spectrometer with the best forecasting effect as a main instruments and the other spectrometers as auxiliary instruments; and V. selecting optimal samples for all the auxiliary instruments from a correction set of the main instrument, determining samples in conversion sets of all the auxiliary instruments according to the sequence numbers, calculating regression coefficients through the unary linear regression, correcting spectrums of the auxiliary instruments, and substituting the regression coefficients into the correction model of the main instrument to obtain a sample component content result. According to the method, the difference of the different instruments is effectively eliminated; the correction model constructed on the main instrument can be shared on the multiple auxiliary instruments; the analysis test workload is reduced, and the cost of the model construction is saved; furthermore, the number of calculated parameters is small, the models are simple, and the forecasting accuracy is relatively high.

The invention discloses a geometric measurement method of Poisson's ratio and Young's elastic modulus of a thin film material, the periphery of the thin film material to be detected is clamped for leading the thin film material to form a circular thin film with the radius of a and the clamped periphery, evenly distributed loads q are imposed on the circular thin film, the deflection value w(r) ofany two points on the circular thin film is measured, a calculation expression formula of the Poisson's ratio and the Young's elastic modulus of the thin film is deducted according to the accurate analytical solutions which are given out by Hencky of the circular thin film with the clamped periphery under the action of the evenly distributed loads, and the Poisson's ratio and the Young's elastic modulus of the thin film can be accurately calculated by utilizing the deflection value ratio of any two points on the circular thin film. The geometric measurement method of the Poisson's ratio and the Young's elastic modulus of the thin film has the advantages of simple and feasible operation, few measurement parameters, more accurate solved Poisson's ratio and the Young's elastic modulus and definite mechanical and physical significances.

The invention provides a lane decision method for automatic driving vehicles based on a multi-objective decision matrix. The lane decision method includes the following steps that (1), a left lane, alane where the automatic driving vehicles are located and a right lane of the automatic driving vehicles are regarded as to-be-decided lanes, continuous-running distance and speed-limiting informationof each to-be-decided lane are collected, quantification is carried out on the information of each to-be-decided lane and evaluated information of each to-be-decided lane is obtained; (2), the information of barriers, closest relative to the vertical distance of the automatic driving vehicles, of the front side and the back side of each to-be-decided lane is collected, current location and speedinformation of the two barriers on the front and back and the location information after setting time are judged, and the relative distance the two barriers on the front and back relative to the automatic driving vehicles correspondingly at the current time and setting time is calculated; (3), integral evaluating information of the to-be-decided lanes is obtained by arranging the above informationand the decision matrix is formed; and (4), a decision lane is obtained by dealing with and calculating the decision matrix, so that the real-time independent lane deciding of the automatic driving vehicles is achieved.

The invention provides a document-level sentiment analysis method based on specific domain sentiment words. The method is implemented by the following steps of collecting a document data set, traininga set of prototype words by using a Skip-gram word vector model to obtain a word vector corresponding to each prototype word, recombining the word vectors by utilizing an attention mechanism, and capturing a relation between non-continuous words in the word vectors; synthesizing the words and sentences by using an asymmetric convolutional neural network and a bidirectional gate recurrent neural network based on the attention mechanism respectively, thereby forming document vector characteristics; generating sentiment eigenvectors by utilizing a domain sentiment dictionary of the Skip-gram word vector model; and finally, combining the document vector characteristics and the sentiment eigenvectors by utilizing a linear combination layer to form document characteristics beneficial to document classification. The sentiment analysis is widely applied to the product analysis, the commodity recommendation, the stock price trend prediction and the like; and the method provided by the invention can accurately and efficiently carry out sentiment analysis on documents, and has great commercial values.

The invention relates to an improved auto-disturbance rejection controlling method, and provides the improved auto-disturbance rejection controlling method to overcome the defect that a nonlinear auto-disturbance rejection controlling method has a plurality of parameters which are difficult to set. According to the improved auto-disturbance rejection controlling method, an extended state observer in a traditional auto-disturbance rejection controlling method is designed as a disturbance observer which is not based on a system model; the disturbance observer is designed to be only related to two parameters by using a linear auto-disturbance rejection controller parameter setting method, so as to achieve the purpose of reducing the amount of setting parameters. Moreover, the improved auto-disturbance rejection controlling method fully integrates the advantages and disadvantages of linear auto-disturbance rejection and nonlinear auto-disturbance rejection, so that the performance of the controller is maximized; the controlling method can reduce the error of the traditional auto-disturbance rejection controlling method, and improves the control accuracy of the system.

The invention discloses a quick text detection method for a natural scene, and relates to the field of image processing, in particular to a novel application technology of a convolutional neural network for text detection. The method is characterized in that a quick feature extraction small network is proposed, an inception module is used, and a small convolution kernel is adopted, so that the number of parameters is reduced, the network is reduced, and the operation speed is accelerated; by additionally arranging a deconvolution layer to integrate multi-scale information, the detection accuracy is improved; a detection framework based on SSD (single shot multibox detector) is adopted in the detection phase, the width-to-height ratio of a preset framework is improved, and the ratio suitable for the text features is adopted. The method provided by the invention has the advantages that the testing is performed on a public dataset, and the validity and real-time property of the method areverified.

The present invention relates to an automatic exposure regulation method and its control system. Said method mainly includes the following steps: setting exposure control basic parameters including target luminance standard value Ytarget, luminance threshold value Ythd, luminance gain value Gtarget, and luminance gain value threshold value, making exposure, extracting image and making image processing, reading exposure time T1, luminance gain G1 and image exposure intensity statistical value Y1, judging exposure intensity value, calculating optimum exposure time T2 and using it to regulate exposure time T1 of image data input module.

The invention relates to the technical field of image processing, more specifically, to a depth convolution neural network image denoising method based on an Inception model. Building a training network with one-layer Inception model; setting the parameters of network solution and choosing the appropriate supervisory framework; the parameters of the network structure are optimized, and the obtained network is used for denoising. The depth convolution neural network image denoising method based on the Inception model of the invention is based on the depth convolution neural network and combinesGaussian white noise to simulate unknown real noise and the network containing the Inception model to reduce the parameter quantity on the basis of maintaining the original denoising effect; throughthe way of data expansion, the content of training data is enriched, and the neural network can learn the inner structure of image distribution or noise. By introducing the residual structure, the network converges more easily and the effect is the best.

The invention provides a mixed recommendation method based on a factorization condition limitation Boltzmann machine. The method includes: constructing a target user feature vector; building a target user grading matrix set; building a target user training sample; building a factorization condition limitation Boltzmann machine model; grading and predicting to recommend results to a target user. The problems of user interest expression inaccuracy caused by pure recommendation based on content and data sparsity caused by recommendation based on behavior are effectively solved.

The invention discloses a flood forecasting method based on depth learning model and BP neural network correction, including: 1.normalizing the historical flood process data, 2. analyzing the normalized historical flood process data to obtain the influence time of each monitoring point on the outlet flow; 3. establishing the input and output values of the forecasting model from the historical flood data by using the sliding window, and establishing the training data set TRSet1 of the model, 4. using TRSet1 to train and establish the flood forecasting model CNNFM; 5, establishing a real-time error correction model training data set TRSet2; 6. training and establishing BP neural network based error correction model BPCM by TRSet2, 7. forecasting the real-time data with CNNFM and correcting them with BPCM to get the final forecast value for the real-time monitoring flow and rainfall. The invention models the training data through the characteristics of automatic extraction of data features by depth learning, and carries out real-time correction through BP neural network to improve the accuracy of model prediction.

The invention discloses a three-dimensional gesture attitude estimation method, comprising: S1) obtaining a plurality of gesture depth images and performing gesture foreground and background segmentation to obtain a plurality of gesture foreground images and to randomly divide the foreground images into a training set and a testing set; S2) according to the gesture model graphs, creating the actual label graphs for the gesture foreground images wherein the actual label graphs contain a plurality of coordinate points representing the various reference identification points of a person's hands in the gesture foreground images with each coordinate point value comprising the coordinate value and the depth value of the corresponding reference identification points; S3) selecting a plurality of gesture foreground images from the training set to train the full-convolution neural network and outputting a plurality of prediction label graphs correspondingly; S4) comparing the deviations between the actual label graphs and the prediction label graphs and at the same time, updating the network parameters; S5) iterating through constantly reducing the deviations until the network parameters are converged; and S6) outputting the to-be-estimated gesture depth graphs and outputting corresponding label graphs as the estimation result.