44results about How to "Good training results" patented technology

The invention, which relates to the technical field of an intelligent vehicle automatic driving and traffic simulation, relates to an artificial intelligent training platform for an intelligent networking vehicle plan decision-making module and aims at improving the intelligent level of the intelligent vehicle plan decision-making module based on enriched and vivid traffic scenes. The artificial intelligent training platform comprises a simulation environment layer, a data transmission layer, and a plan decision-making layer. The simulation environment layer is used for generating a true traffic scene based on a traffic simulation module and simulating sensing and reaction situations to the environment by an intelligent vehicle, thereby realizing multi-scene loading. The plan decision-making layer outputs a decision-making behavior of the intelligent vehicle by using environment sensing information as an input based on a deep reinforcement learning algorithm, thereby realizing training optimization of network parameters. And the data transmission layer connects the traffic environment module with a deep reinforcement learning frame based on a TCP / IP protocol to realize transmission of sensing information and vehicle control information between the simulated environment layer and the plan decision-making layer.

The invention provides an electroencephalogram sleep staging method based on a deep convolutional neural network. The electroencephalogram sleep staging method comprises the following steps that S1, sleep signals of a subject are collected, and multi-lead electroencephalogram signals in the sleep signals are extracted; S2, performing data preprocessing on the electroencephalogram signals; S3, constructing and training an end-to-end deep convolutional neural network classifier; and S4, performing electroencephalogram sleep staging by using the deep convolutional neural network classifier. Compared with the conventional CNN electroencephalogram sleep staging method, the electroencephalogram sleep staging method provided by the invention has the advantages that under the condition of the sameiteration times and learning rate, each batch of the model adopts higher data, and the obtained output result is more stable. In terms of accuracy and F score, the method provided by the invention has better classification performance.

The invention discloses a general steganography method based on deep learning. The method comprises the following steps: S1, carrying out hiding processing on a sender; dividing secret information tobe hidden into n groups of information fragments, wherein each group of information fragments correspond to one category label, a deep learning model is adopted, the category label and random noise are used as drive, a pseudo-natural image of a specified category is generated, and the pseudo-natural image is used as a secret-containing image input channel after hiding processing; and S2, carryingout extraction processing at a receiver: inputting the secret-containing image into a discriminator by the receiver to carry out image authenticity identification and image category judgment, then sending the image category information into a function converter to be processed to obtain a secret information fragment, and decoding the secret information fragment to obtain original secret information. According to the invention, the security and confidentiality of information transmission can be greatly improved.

The invention discloses a welding defect real-time detection method and system based on high-frequency time sequence data. The detection method comprises the steps of firstly sampling the collected high-frequency welding time sequence data according to a set window length, marking a defect occurrence time period and a defect type for each sample, and generating a data sample set; training a ResNet and TCN fusion network model by using the generated data sample set to obtain a trained detection model; and finally, obtaining new real-time high-frequency welding data, inputting the new real-time high-frequency welding data into the trained detection model for prediction according to a set window length, and outputting a welding defect category in real time. According to the method, the ResNet network and the TCN are subjected to network structure fusion, the ResNet can be applied to the field of time sequence detection, and for high-frequency welding time sequence data with a large data size and a long sequence length, the training speed is increased in the training process through a parallel convolution calculation mode, and strong real-time prediction is achieved in the prediction process.

The invention discloses a multi-modal fusion saliency detection method based on a convolution attention module. The method comprises: in a training stage, constructing a convolutional neural network;inputting the left viewpoint image and the depth image of the original image into a convolutional neural network for training to obtain a corresponding saliency detection image; calculating a loss function between a set formed by saliency detection images generated by the model and a set formed by corresponding real human eye gaze images to obtain an optimal weight vector and an offset term of theconvolutional neural network classification training model; and in a test stage, inputting the three-dimensional image in the selected data set into the trained convolutional neural network model toobtain a saliency detection image. According to the visual saliency detection method, extraction of image features is optimized by applying a novel module, multi-scale and multi-mode feature fusion iscarried out, and finally the detection efficiency and the detection accuracy of visual saliency detection are improved.

The invention relates to the technical field of natural language processing, in particular to a question-answering method for entity relationship extraction based on transfer learning. The acquisitionof a relationship classification result comprises the steps: obtaining and preprocessing a source domain text data set and a target domain text data set; inputting the preprocessed data into a skip-gram model for training to obtain word vectors of the source domain text data and the target domain text data, obtaining position vectors of the source domain text data and the target domain text data,and cascading the position vectors with the word vectors to obtain joint feature vectors of the source domain text data and the target domain text data; inputting the joint feature vector of the source domain text data into a BiLSTM network for pre-training to obtain network parameters in the pre-training process and context information and semantic features of the source domain text data; and inputting the joint feature vector of the target domain text data into a BiLSTM _ CNN fusion model for retraining to obtain a high-dimensional feature vector of the target domain text data, sending thehigh-dimensional feature vector into a classifier, and outputting the relationship classification result. Question-answering accuracy can be improved.

The invention discloses a method for testing fault modes of an integrated switching current circuit. The method includes the following steps: S1, establishing a switching current test circuit, and applying an excitation signal to the test circuit, S2, defining the fault modes of the switching current circuit, S3, collecting testable node time-domain response signals of the switching current circuit, S4, preprocessing time-domain response data, calculating fault characteristic parameters of the signals, extracting information entropy and kurtosis of the signals, and calculating a fuzzy set of the information entropy, and S5, constructing a neural network classifier according to the information entropy and the kurtosis, and acquiring the fault modes of the switching current test circuit. The method for testing the fault modes of the integrated switching current circuit is suitable for tests on a large-scale and complicated integrated switching current circuit with a large number of fault categories and is high in accuracy rate of fault diagnosis.

The invention discloses a medical image focus detection modeling method, device and system based on federated learning, and the method comprises the steps: enabling a global server S to transmit a generated global parameter [omega]<0> to local focus recognition clients C<k>; generating a global parameter [omega]<theta+1> by using detection head network parameters returned by the K local focus identification clients C<k>; and sending the global parameter [omega]<theta+1> to each local focus identification client C<k> to obtain a corresponding medical image focus detection model. The training information of the intermediate model is transmitted to a data holder through codes, respective data information does not need to be shared, and the model is integrated through a corresponding strategy, so that better training and prediction results are returned.

The invention relates to a training device, in particular to a medical leg rehabilitation training device. The to-be-solved technical problem is to provide a medical leg rehabilitation training devicewith a high safety factor and a significant training result. In order to solve the above technical problem, the provided medical leg rehabilitation training device comprises a mounting plate, a firsthollow rod, a first connecting rod, a height adjusting mechanism, a bearing plate, a first fixing plate, a second fixing plate, a rotating mechanism, a second hollow rod, a first connecting block, athird hollow rod, a U-shaped handrailing, a fixing mechanism, a resetting mechanism and a pedal; the first hollow rod and the second fixing plate are fixedly connected to the top of the mounting platein sequence in the length direction of the mounting plate, and the first connecting rod is slidably connected with the first hollow rod. The medical leg rehabilitation training device achieves the effects of the high safety factor and significant training result.

The invention belongs to the technical field of artificial intelligence, and discloses an OCT image denoising method and device based on an annular generative adversarial network, and the method comprises the steps: obtaining a to-be-denoised OCT image; inputting the to-be-denoised OCT image into a trained annular adversarial generative network model; and outputting a denoised OCT image through the annular adversarial generative network model. According to the method, the OCT image is denoised through the annular adversarial generative network model, and the high-noise OCT image is effectivelyconverted into the clear OCT image, so that a doctor can read the image or use the OCT image for software analysis. Moreover, the limitation that training data must be paired in denoising applicationin previous deep learning is avoided, and acquisition of a large amount of data for training is facilitated, so that the denoising effect of the model is improved.

The invention discloses a rapid ship target detection method, a storage medium and computing equipment, and the method comprises the steps: establishing a feature pyramid full-convolution network anda double-branch module which sequentially comprise an input layer, a feature extraction layer, a feature fusion layer and an output layer, determining a ship data set, inputting a generated training set into the built feature pyramid full-convolution network, and generating a target detection result. According to the method, the ship in the image can be rapidly and accurately detected; and the detection result is accurate and rapid, the requirement on embedded equipment is low, and the method has very high practical application value in various aspects such as military affairs, civil use and the like.

The invention provides an air conditioner on-hook spatial distribution automatic identification method and system based on a streetscape picture, which realizes air conditioner on-hook spatial distribution identification by taking the streetscape picture provided by a network platform as a data source, and comprises the following steps: capturing the streetscape picture through a crawler; and pre-screening the crawled street scene pictures based on the air conditioner hanging scene characteristics, labeling a window entity and an air conditioner hanging target entity, inputting the window entity and the air conditioner hanging target entity into a YOLOX network for air conditioner hanging entity model training and window entity model training, and after training is completed, predicting each street scene picture by using a corresponding model. Outputting the number of air conditioner outdoor units and window entities in each picture; and calculating the ratio of the number of air conditioner outdoor units to the number of windows, performing space projection in combination with the coordinate information of each street view picture, and outputting the corresponding air conditioner coverage rate of the region. By mining the network streetscape picture information, identifying the ratio of the air conditioner hanging unit to the window and extracting the air conditioner space distribution information, the labor cost can be greatly saved.

The invention discloses an infant abnormal behavior detection method based on a meanshift algorithm and an SVM. The method comprises the following steps: preprocessing an acquired baby video; performing target motion trail tracking on four limbs and the whole body of the baby in the video by using a meanshift algorithm; storing the obtained motion trail information; then, wavelet transform is usedfor extracting motion trail information; establishing a sample set for the extracted wavelet approximate waveform; the set SVM support vector machine is used for training; solving a power spectrum ofthe motion trail information by using wavelets; and establishing a sample set by using the obtained features, training the sample set by using the set SVM support vector machine, testing the two trained models, and setting different weight parameters by using a data weighted fusion algorithm according to different accuracies of the two models so as to carry out weighted judgment, thereby obtaining an optimal training result.

The invention belongs to the technical field of artificial intelligence assisted medical examination, and discloses a malformed urine erythrocyte classification statistics method and system. The method comprises the following steps: S1, collecting a microscope zoom video of a sample; S2, identifying abnormal urine erythrocytes on the frame with the highest definition in all frames of the microscope zoom video, and segmenting a plurality of abnormal urine erythrocyte zoom videos; S3, predicting the probability of each frame in each deformed urine erythrocyte zoom video under different classifications by utilizing a deep multi-example learning algorithm; and S4, realizing classification and quantity statistics of deformed urine erythrocytes through a target-shaped key frame priority principle. The method can accurately detect, classify and count the abnormal urine red blood cells in the sample, is accurate and reliable, and can be applied to the field of urine sample detection.

The invention discloses a resolution-by-resolution improved image super-resolution restoration method based on an attention mechanism, which is mainly based on a resolution-by-resolution improved super-resolution restoration network of a double attention mechanism, and improves the feature extraction capability of a model by introducing a convolution module of a feature dimension attention mechanism and a spatial dimension attention mechanism. Furthermore, related ideas of a resolution-by-resolution improvement network in the field of adversarial neural networks are used for reference, the network learning difficulty is simplified, and super-resolution restoration of resolution-by-resolution improvement is realized. Finally, the algorithm performance is tested through a DIV2K data set, andexperimental results show that the method can improve the input low-resolution image resolution by resolution, the same network can improve the resolution by 2-4 times at the same time, and the PSNRvalue of the reconstructed image is significantly superior to that of a current mainstream algorithm.

The invention discloses a method and a system for improving seismic data resolution based on a weak supervision generative adversarial network. The method comprises the following steps of: performing normalization processing on two pieces of three-dimensional seismic data belonging to different work areas; dividing a training set and a test set; obtaining a training sample pair in the training area in a random extraction mode; sending the low-resolution seismic data into a forward generator; sending the output of the forward generator to a reverse generator; sending the output of the forward generator and the corresponding high-resolution label to a discriminator for discrimination; and alternately training the generator and the discriminator, continuously updating network parameters until the model converges, and after training is finished, sending the whole block of low-resolution seismic data into the forward generator for testing to obtain a final high-resolution result. According to the method, the distribution characteristics of the high-resolution seismic data can be learned on the premise that paired input and labels do not exist, and the high-frequency information of the original seismic data can be accurately and effectively recovered.

The invention provides a target detection method and device, and the method comprises the steps: adding a pseudo tag for label-free data, and dividing the pseudo tag into a high-quality pseudo tag and an uncertain pseudo tag; inputting the label-free data into the initial learning model to obtain a first predicted value; determining a first prediction label and a first prediction frame based on a first prediction value corresponding to the high-quality pseudo label, and determining a second prediction label and a second prediction frame based on a first prediction value corresponding to the uncertain pseudo label; inputting the label-free data into the initial management model to obtain a second prediction value, and determining a third prediction label and a third prediction frame based on the second prediction value corresponding to the uncertain pseudo label; and training the initial management model based on the first prediction label, the first prediction frame, the second prediction label, the second prediction frame, the third prediction label and the third prediction frame to obtain a target management model which is used for performing target detection on the to-be-detected data. Through the scheme of the invention, acquisition of a large amount of labeled data is avoided.

The invention discloses a push-up assisting device suitable for beginners, which comprises a trunk supporting plate, wherein the trunk supporting plate is provided with a head end close to the head ofa user and a tail end close to the foot of the user; a foot supporting plate rotatably connected to the tail end of the trunk supporting plate and arranged in a vertical direction; an auxiliary rod wherein one end of the auxiliary rod is rotatably connected to the lower surface of the trunk supporting plate, and the other end of the auxiliary rod slidably braces against the ground; an elastic piece is also arranged between the auxiliary rod and the trunk supporting plate, and the elastic piece keeps the trunk supporting plate and the auxiliary rod at a certain angle. The push-up assisting device suitable for beginners can relieve upper limb pressure, is conducive to standard actions completion, reduces injury risks, and is simple in structure and convenient for use.

The invention discloses a natural language processing model training method and device, and relates to the technical field of deep learning and natural language processing. The specific implementationscheme is as follows: generating derivative models according to an obtained meta-model set of natural language processing, and adding the plurality of derivative models into the meta-model set as meta-models to increase the number of models in the meta-model set for subsequent meta-training of the meta-model set; and then, screening the meta-models in the meta-model set according to the performance parameters of the trained meta-models to obtain the meta-models with good performance for adaptive training of natural language processing tasks. Due to the fact that the scheme adopts the mode that the meta-models are enriched and then screened, the performance of the screened and reserved meta-models is improved, and no matter what fields or languages are involved in adaptive training, accurate processing results can be obtained on subsequent natural language processing tasks of the corresponding fields or languages.

The invention discloses a model accelerated training method and device based on training data similarity aggregation, and the method comprises the steps: taking a part of minimized training data as a starting point, extracting data with poor prediction from a prediction result of a current model in a mode of random sampling and random increment in each round, and sampling additional training data in a clustering extraction mode, therefore, the most representative training information is obtained, and the training efficiency of each round is improved. The data set scale of each round of model training is reduced, the training time is greatly shortened, clustering does not need an accurate result, the number of iterations can be reduced or a faster and simpler clustering method is used, and the total training time of each round is still much shorter than that of original full training set training on the whole; the training data selected in each round is targeted, the images with inference errors are selected for training, the back propagation gradient can be obtained to the maximum extent, the probability of falling into the local optimal solution during training is reduced, dynamic adjustment in the training process is facilitated, and the optimal training result is achieved.

The invention provides a multilingual end-to-end OCR algorithm and a multilingual end-to-end OCR system, which overcome and bypass the defects of fragments in the prior art and have excellent performance on character adhesion, Chinese-English and median mixed data. By means of a self-distillation transformer module, the position relation is reserved, parameters and model complexity are reduced, meanwhile, results are output in parallel, dependence between nodes is cut off, higher robustness is achieved for multilingual and multi-font scenes, and the structure and performance are optimized. Thealgorithm comprises the steps of obtaining a feature map of a to-be-identified picture; training the feature map through a relation attention module based on the self-distillation transformer moduleto obtain a character matrix; performing parallel attention decoding on the character matrix to obtain a prediction result; and according to the prediction result, based on the vocabulary sentence table, obtaining an OCR model conforming to the language of the vocabulary sentence table.

The present invention relates to a craniofacial contour surgery simulation teaching aid. The craniofacial contour surgery simulation teaching aid comprises a head mold for simulating a human head and an elastic head hood; the elastic head hood covers the head mold; and a first opening corresponding to the dentition of the head mold is formed in the elastic head hood. With the craniofacial contour surgery simulation teaching aid of the invention adopted, an intraoral approach craniofacial contour surgery can be reproduced vividly and realistically; and the first opening is formed in the elastic head hood, so that a professional doctor can carry out the training of the intraoral approach craniofacial contour surgery, time for a plastic surgery doctor to master the craniofacial contour surgery can be shortened, and a surgery training effect can be ensured.

The invention belongs to the technical field of face recognition, and discloses an artificial intelligence recognition device based on big data, and the device comprises: a three-dimensional laser scanning module which obtains face information data through a three-dimensional low-light scanner, and meanwhile, preprocesses the three-dimensional face point cloud; the three-dimensional image preprocessing module that is used for extracting three-dimensional multimedia visual target image features by analyzing two aspects of visual two-dimensional multimedia target image feature extraction and three-dimensional multimedia visual target image change feature generation; the face model training module that is used for calling an OpenCV internal function to obtain a trained model based on an OpenCV technology; the interaction recognition module that is used for inputting face information data and carrying out recognition; and the remote face recognition module that performs data processing and face recognition in the LabVIEW development environment, and finally feeds back the recognized identity information to the slave computer. According to the invention, the computer technology can be developed to intelligence and automation.

The invention discloses an SVM-based credit default prediction method under differential privacy in the technical field of credit default. The SVM-based credit default prediction method comprises thefollowing steps of S1, preprocessing data; S2, carrying out variable selection; and S3, designing a weighted SVM optimization model under differential privacy according to a differential privacy serial combinatorial property. An effective solution is provided for a differential privacy SVM learning problem under data imbalance, and the solution can specifically solve the data imbalance problem faced when differential privacy SVM learning is used to predict customer default. The method is suitable for application scenes with data imbalance including credit card default prediction, such as disaster prediction, medical diagnosis and other fields, and is also suitable for the technical scheme of the invention.

The invention provides a construction method of a neural network architecture for simulating dendritic spine change, which comprises the following steps: simulating a dendritic spine in the brain of a higher animal during birth by using a neural network, storing the weight of the neural network by using an adjacent matrix, and generating a weight matrix; initializing the weight matrix, simulating the pruning process of dendritic spines in the brain during growth and development of higher animals, and genrating the initialized weight matrix; obtaining training samples, and dividing the training samples into a plurality of groups, wherein each group comprises the same number of training samples; inputting each of the plurality of groups into the initialized weight matrix for training, simulating the learning process of higher animals, and generating a trained weight matrix; and converting the trained weight matrix into a real network architecture, wherein the real network architecture represents the dendritic spines of the brain of the higher animal after learning. The method can be used for a supervised image recognition task, can train a proper neural network architecture for different problems, and has high adaptability.

The invention discloses a model training method and device, electronic equipment and a storage medium, and relates to the technical field of deep learning. According to the specific implementation scheme, derivative models are generated according to an obtained meta-model set, the multiple derivative models are added into the meta-model set as meta-models, and the number of models in the meta-model set is increased for subsequent meta-training on the meta-model set; and then, the meta-models in the meta-model set are screened according to the performance parameters of the trained meta-models so as to obtain the meta-models with good performance for adaptive training of a target task. According to the model training method and device, the electronic equipment and the storage medium, the manner of carrying out screening after enriching the meta-models is adopted, so that the performance of the screened and reserved meta-models is improved, an adaptive training result does not need to bereversely transmitted to a meta-training process, a better training result can also be achieved, and the training efficiency is effectively improved.