73 results about "Continuous training" patented technology

The invention relates to a binocular depth estimation method based on a depth neural network, which comprises the following steps: 1) preprocessing the input left and right viewpoint images to enhancedata; 2) constructing a multi-scale network model for binocular depth estimation, wherein the model comprises a plurality of convolution layers, an activation layer, a residual connection, a multi-scale pooling connection and a linear upsampling layer; 3) Designing the loss function to minimize the results in the continuous training process, so as to obtain the optimal network weights; 4) inputting the image to be processed into the network model to obtain the corresponding depth map, and repeatedly repeating the above steps until the network converges or reaches the training times. The invention adopts the idea of unsupervised learning, and only the left and right viewpoint images obtained by the binocular camera are used as network input. The adaptive design of the network sets the internal and external parameters of the camera as a single model parameter, so it can be applied to multiple camera systems without modifying the network.

The invention discloses an LSTM network-based treatment and convalesce combined health monitoring method. The method comprises the following steps that by means of health detection equipment, intelligent home and an HIS system, comprehensive health information data of people aged 60 years and above is collected and obtained and then stored in a cloud database; collected and obtained data waveformsare subjected to denoising processing by adopting an adaptive notch filter of LMS; processed and classified data serves as a training sample, a pervasive health model of people aged 60 years and above is generated by adopting LSTM network training and learning; for an existing pervasive health model, after personal data is obtained, continuous training and personalized adjustment are conducted while the model is identified, and abnormal early warning of the corresponding situation is given. Accordingly, a traditional statistical model algorithm and a neural network are organically combined, comprehensive health monitoring is performed on people aged 60 years or above, efficient and accurate personalized health prediction and early warning are achieved, and a solid foundation is laid for building intelligent health pension service.

The invention discloses a deep learning palm vein recognition system based on cloud side cooperative computing. The system is mainly composed of a cloud computing layer, an edge computing layer and aterminal layer. The calculation and storage capabilities of the whole identification system are determined through the combination of three layers; the system specifically comprises a cloud storage module, palm vein acquisition and identification equipment and a client; the cloud storage module can provide computing resources required by training for the model and transmits the trained model to anedge calculation module of the palm vein acquisition equipment; the terminal collects the recognition condition of the deep convolutional neural network model, and feeds back the recognition condition to the deep convolutional neural network model chip on the cloud storage module through the edge calculation module.Continuous training, application, feedback and retraining are performed, and the accuracy of the finally obtained palm vein recognition model is greatly improved through repeated effective iterative training. The rear-end cloud storage module can also realize rapid deployment and unified management of the palm vein recognition system according to the resource configuration and management capability of the cloud platform.

The invention, which belongs to the field of the deep learning neural network, discloses a multi-feature-learning-based adversarial network training method. The method comprises the following steps: S1, constructing a deep convolutional generative adversarial network (DCGAN) model; S2, constructing a plurality of convolution kernels for a determination device; S3, initializing random noises and inputting the processed noises into a generator; S4, with the plurality of constructed convolution kernels with different sizes, carrying out conversion processing on all images received by the determination device to obtain a plurality of characteristic patterns; and S5, outputting a mean value of loss functions of the plurality of characteristic patterns into the generator for continuous training.According to the method disclosed by the invention, the training method after picture receiving by the determination device is changed and thus the convolution process of the determination device iscarried out in parallel, so that a plurality of characteristics of the images generated by the generator are learned simultaneously and several kinds of characteristics in a data set are learned rapidly and thus images matching the characteristics of the data set are generated efficiently.

A pedestrian attribute recognition method based on image and attribute joint guidance includes: obtaining image features and attribute features; build an image-attribute mutual guidance mechanism, including constructing the features of image guiding attributes and pedestrian attributes guiding images; inputting the feature of image guide attribute and the feature of pedestrian attribute guide image into a short-and long-term memory model, and obtaining a pedestrian attribute recognition model in which the image and attribute guide each other; obtaining the pedestrian attribute result and optimizing the pedestrian attribute result by continuous training using cross-entropy objective function. The method can make the model learn the relationship between the image feature and the attribute feature better through the mutual guidance between the image feature and the attribute feature, and adds the attention mechanism in the attribute recognition process, so that the model can further improve the distribution of the two features to improve the accuracy of the pedestrian attribute recognition.

Apparatus and method for improving recognition of patterns such as alphanumeric characters. A known recognition system is expanded to further include a complementary recognition system which is linked with the primary recognition system. An image that can not be positively recognized by the primary recognition system is passed on to the complementary recognition system and any characters not positively recognized by the complementary recognition are again passed on to a correction system. At the correction system, an operator classifies unrecognized characters which are then used to teach the complementary recognition system. Thus, the classified data of the correction system provide the training data for a continuous training process which is coupled with the correction system by a pattern adaptation system.

The invention provides a method for detecting the characteristics of a video object in a finite complex background by studying color-based and gradient-based background models. The method comprises the following steps: 1, taking a background scene graph, conducting continuous training on a background sample image to obtain a color-based background model of mixed Gaussian distribution. If the background model conforms to the mixed Gaussian distribution, the gradient distribution functions of various combinations need to be calculated, and if a certain pixel belongs to any gradient distribution, the background model is supposed to conform to the gradient background model; 2, calculating the gradient distribution functions of each image to be detected, and establishing a color- and gradient-based background model; and 3, updating the parameters and weight of Gaussian distribution according to the ambient changes, such as illumination, wind strength and the like, and further updating thecolor-based models and the color- and gradient- based models. Compared with the prior art, the invention has the characteristics of background disturbance resistance and adaption to changes in environmental illumination.

The invention particularly relates to a method for efficiently utilizing cloud center quantum computer resources. According to the method for efficiently utilizing the cloud center quantum computer resource, a quantum cloud center creates a task for a quantum computing application program, and after testing, simulating, simulating, evaluating and optimizing the program, the task enters a physicalquantum computer execution program queue to wait for execution; the quantum cloud center comprehensively executes historical data generated by similar quantum computing application programs, a multiplexing result is selected, the quantum cloud center dynamically selects the programs and loads the programs to the physical quantum computer, and the use efficiency of the quantum computer is improved.The invention discloses a method for efficiently utilizing cloud center quantum computer resources, resources can be distributed more reasonably and efficiently, the use efficiency of the physical quantum computer is improved, the quantum program task selection model and the evaluation model can be optimized through continuous training, the rationality of quantum computing application program selection is continuously improved, and then the operation efficiency of the physical quantum computer is improved.

The invention discloses a network protocol test case generation system based on neural style migration, which comprises an original data collection module, an encoding module, a neural style migrationmodule and a reverse encoding module. The original data collection module collects the traffic data in the industrial control network system and classifies the collected data by clustering algorithm.The encoding module encodes the classified data into a picture form; The neural style transfer module takes the pictures and style pictures output by the encoding module as the input, and transformsthe neural style based on the neural style transfer method. In the process of transforming, the degree of the style transformation is controlled by continuous training iteration. The inverse coding module converts the two-dimensional image output by the neural style transformation module into one-dimensional sequential form as a test case. This test case can be directly injected into the target network for attack testing. This system can intelligently learn the protocol format, reduce the artificial learning process, and improve the efficiency of testing.

The invention discloses an asynchronous distributed deep learning training method, device and system. The method is applied to a computing node k, the computing node performs the forward propagation through the local model parameters and the local data to obtain the loss values, then the gradient updating is obtained via the backward propagation, and only the values with larger information amount in the gradient updating vectors are sent by using the sparse function, so that the communication traffic from the computing node to the parameter server is effectively reduced. On the basic concepts of the asynchronous random gradient descent and a parameter server, the gradient of the computing node is returned via the parameter server, the communication compression is carried out, the compression rate of the communication is effectively guaranteed. The compression rate of communication is continuously increased along with the continuous training, and a better compression effect and a higher communication speed can be achieved.

The invention relates to the technical field of orthopedic training frames, in particular relates to an intermittent leg rehabilitation training frame for the orthopedics department which comprises ahorizontal plate and support legs, the support legs are fixed at four corners of the horizontal plate; two shank sliding grooves are formed in the left end of the upper surface of the horizontal plate. A first wall groove is formed in the outer side wall of the shank sliding groove; a shank sliding block is slidably arranged in the shank sliding groove, a connecting block moving left and right along the first wall groove is arranged on the outer side face of the shank sliding block, a shank containing groove is formed in the upper surface of the shank sliding block, a foot sole limiting plateis arranged at the left end of the shank containing groove, and a groove is formed in the middle section of the upper surface of the horizontal plate; through the meshing effect between a moving frameand an incomplete gear, the legs of a patient are subjected to rehabilitation training in the bent, pause and straightened states, continuous training in the bent and straightened states all the timewhen an existing rehabilitation training device is used for training the legs of the patient is effectively avoided, and the rehabilitation training effect of the rehabilitation training frame is better.

The invention discloses a white balance proofreading method based on a white object, which comprises the following steps: a white object contrast model is established by using a large number of acquired white object samples through continuous training and learning; The white object contrast model is applied to the scene capture, and the pixel mean value and color gain value are calculated by identifying the image and extracting the white object region in the image, so as to adjust the color to realize the white balance proofreading of the image; The invention has the advantages of simple realization principle, low resource cost, fast operation time, ideal and accurate white balance effect. The paper towel identification technology adopted by the invention is based on convolution neural network algorithm, learning from the large sample data, extracting the high-level features of the image gradually, classifying the features and completing the recognition can deal with the offset, scalechange and deformation of the paper towel to a certain extent, ensure the strong separability of the features, have an ideal recognition effect on the feature classification, reduce the dependence ofthe recognition on external conditions, and reduce the complexity of the model.

The invention relates to an ankle-joint rehabilitation training assisting device. Other people need to assist a patient with the dysfunctional ankle joint of the lower limb in carrying out rehabilitation, and the patient cannot complete rehabilitation training independently. The ankle-joint rehabilitation training assisting device is provided with a base which is provided with a rotary pedal, one end of the rotary pedal is fixed to the base through a connection shaft, and the other end of the rotary pedal is fixed to a telescopic adjustable moment part arranged on the base. The rotary pedal is provided with a foot support, guide posts are fixed to the two sides of the rotary pedal, each guide post is fixedly provided with a leg support, and the patient carries out rehabilitation by stamping on the rotary pedal. Through the ankle-joint rehabilitation training assisting device, without the help of other people, the patient can independently control the training time sequence and the training progress at the right moment according to requirements of the patient, the action capability of the joint and muscle force for relevant control can be improved and enhanced through continuous training, and rehabilitation is facilitated.

The invention discloses a method for auditing the content of bad information buffered by CDN and CACHE. The method comprises the following steps of A acquiring a data packet through a data transmission interface; B carrying out the data format of a file analysis module; C using that file analysis module to generate URL to be scanned and transmit the URL to the web page content grabbing module; D using the data processing module to use the intelligent image recognition model to recognize the specific image of the picture, and complete the process of identifying and auditing the CDN/cache content. Compared with the prior art, the invention improves the service coverage range of the content audit, supports the bad information audit of the CDN/webcache service, and the crawler module can directly read the URL to be scanned through the file to carry out content crawling and obtain the web page content. By using the depth learning algorithm and simulating a human brain neural network, a model with high-level expressive force is constructed. Through the continuous training of big data and the frequent iteration, the precision can reach 99.5%.

An embodiment of the invention provides a pilot frequency distribution method and a pilot frequency distribution device of a multi-antenna system. The pilot frequency distribution method comprises the following steps of: determining the number of pilot frequencies to be distributed and the number of training periods according to the number of users of each cell; and determining a pilot frequency distribution scheme of each cell according to the number of pilot frequencies and the number of training periods, wherein determining the pilot frequency distribution scheme of each cell comprises the following steps of: determining the pilot frequency for each training period, which is distributed by each user of lth cell, and distributing the pilot frequency of at least kth user of the lth cell to different users of other cells in the system in each training period. According to the pilot frequency distribution method and the pilot frequency distribution device, which are provided by the embodiment of the invention, the pilot frequencies distributed to the expected users are distributed to different users in each training period, so that the pilot pollution in different continuous training periods is randomized for providing a foundation for carrying out later accurate channel estimation.

The invention provides an environmental protection intelligent law enforcement emergency traceability scheduling management method based on big data. The environmental protection intelligent law enforcement emergency traceability scheduling management method based on big data includes the steps: 1) establishing a customized triangle environment processing efficiency acquisition and assessment model for a law enforcement object, the law enforcement time and a law enforcement officer; 2) performing self-adaption layered classified dimensionality reduction on the air pollution data; and 3) by means of the customized triangle environment processing efficiency acquisition and assessment model for the law enforcement object, the law enforcement time and the law enforcement officer, according tothe air pollution data after self-adaption layered classified dimensionality reduction, performing continuous training subdivision by combining with the big data ant colony algorithm.

The invention discloses a PDF document content text paragraph aggregation method based on a neural network, and the method comprises the steps: defining dozens of features of a row of texts, converting the features into multi-dimensional vectors, generating a sample data set, designing an algorithm model, carrying out the continuous training of the model, and finally outputting the trained algorithm model. For two input lines of texts, the algorithm model is used to accurately determine whether the two lines of texts should be merged into the same paragraph. Based on an artificial intelligencetechnology of a neural network, a research and development application program automatically aggregates line characters extracted from PDF into paragraphs, original sentences and paragraph structureinformation of the characters are restored, and repeated utilization of PDF content data is facilitated. The automatic aggregation efficiency of the artificial intelligence program cannot be achievedthrough manual processing, manual work is replaced by machines, the labor cost is saved, and the efficiency is greatly improved.

The invention relates to the technical field of fitness equipment, provides an outdoor sports comprehensive warm-up fitness training all-in-one machine, and aims to solve the problems that a common outdoor training device is low in applicability and cannot perform comprehensive training according to requirements of users, and during training, a trainee cannot have continuous training power and desire. The all-in-one machine comprises a horizontally-arranged bottom plate and a machine body vertical shell arranged on the bottom plate. A machine body top shell is mounted at the top of the machinebody vertical shell in a penetrating manner, and the machine is characterized in that a transverse mounting part extends from one side of the lower part of the machine body vertical shell, a leg exercising mechanism is arranged in the transverse mounting part, a cushion is mounted at the top end of the transverse mounting part through an adjusting frame, and a waist massaging mechanism is arranged in the middle of the side wall of the machine body vertical shell; and the leg exercising mechanism comprises a counterweight flywheel which is rotatably inserted into the transverse mounting part.The all-in-one machine is particularly suitable for outdoor sports comprehensive training and has high social use value and application prospects.

The invention discloses a network protocol test case generation method based on neural style migration, which comprises a raw data collection step, an encoding step, a neural style migration step anda reverse encoding step. The original data collection step collects the traffic data in the industrial control network system and classifies the collected data by clustering algorithm. The encoding step encodes the classified data into a picture form. The neural style transfer step takes the encoded picture and the style picture as the input, and transforms the neural style based on the neural style transfer method. The degree of the style transfer is controlled by continuous training iteration in the process of transforming. The inverse coding step converts the two-dimensional image output from the neural style transformation step into one-dimensional sequential form as a test case. The test case can be directly injected into the target network for attack testing. The method can intelligently learn the protocol format, reduce the artificial learning process, and improve the efficiency of testing.

The invention discloses a marine noise signal identification method based on a deep neural network (DNN). According to the method, a deep-neural-network model is established, continuous training and updating of forward operating and back propagation are carried out on weight values of each layer of neurons in the model to obtain classification weight values capable of distinguishing different types of ocean noise signals, and thus identification on the different types of ocean noise signals is realized. According to the identification method of the invention, a deep belief network is utilizedto carry out initial weight value training of the deep neural network, obtained weight values are used as initial weight values of training of the deep neural network, then training is carried out ondata, thus identification on the different types of ocean noise signals is realized. According to the method, the deep neural network and the initial values trained by the deep belief network are utilized to enable accuracy of test results to be high, and high-precision identification requirements can be met.

The invention discloses a difficulty-adjustable dynamic balance ability training instrument, particularly relates to an active and difficulty-adjustable human body dynamic balance ability training instrument, and belongs to the field of training instruments. The instrument is mainly composed of a training pedal, a cylindrical pin, a hollow hemispheroid, a base, a rope, a bush, a guide pipe, a supporting beam and an adjustable hydraulic damper. The adjustable hydraulic damper comprises a plunger, a plunger sleeve, a spring, an adjusting knob, hydraulic oil, a sealing ring and a check valve. According to the invention, a user can select the corresponding training difficulty according to the balance ability level of the user, the dynamic balance ability of the user can be gradually improved through continuous training, and related sport injuries can be avoided. An inertia sensor used for measuring the dynamic balance ability of the user is installed on a training pedal of the dynamic balance ability training instrument, and dynamic balance ability data of the user are measured through the inertia sensor. The device has the advantages of being low in cost, portable and easy to popularize on a large scale.