24 results about "Deep neural networks" patented technology

A method and device for voiceprint recognition, include: establishing a first-level Deep Neural Network (DNN) model based on unlabeled speech data, the unlabeled speech data containing no speaker labels and the first-level DNN model specifying a plurality of basic voiceprint features for the unlabeled speech data; obtaining a plurality of high-level voiceprint features by tuning the first-level DNN model based on labeled speech data, the labeled speech data containing speech samples with respective speaker labels, and the tuning producing a second-level DNN model specifying the plurality of high-level voiceprint features; based on the second-level DNN model, registering a respective high-level voiceprint feature sequence for a user based on a registration speech sample received from the user; and performing speaker verification for the user based on the respective high-level voiceprint feature sequence registered for the user.

The invention relates to a deep neural network target detection method integrated with a pixel-level attention mechanism. The deep neural network target detection method comprises the following steps:collecting a training sample image; converting data in the collected image data set into a format required by training the deep convolutional neural network through preprocessing, and performing dataenhancement to improve the training effect of the neural network; designing a deep convolutional neural network structure integrated with a pixel-level attention mechanism, the network structure being used for receiving an input image and outputting bounding box regression information and category information to an object in the image: the network structure comprising two parts, one part being abasic network for preliminarily extracting features; the other part being a plurality of cascaded convolutional layers or downsampling layers added at the tail end of the basic network and being usedfor extracting convolutional feature maps with different scales and fusing the multi-scale feature maps with a pixel-level attention mechanism; and training.

The invention discloses an odor identification method based on gas sensors and deep learning, which is characterized in that a response curve cluster of odor to be detected is obtained through a gas sensor array, original data is directly used as an input sample of an odor identification deep neural network, data preprocessing and data amplification are performed on the input sample, time sequenceresponse data hierarchical features are automatically extracted by adoption of deep learning, global feature extraction and long-range dynamic feature extraction are performed at the same time, and an odor label is output through a classifier, so that high-sensitivity and specific odor identification is realized. The odor identification method of the invention has high sensitivity and high reliability, and can be widely applied to the fields of industrial production, medical treatment, environment, safety and the like.

The invention discloses an industrial part key point detection method based on deep learning in the field of industrial vision. The method specifically comprises the following steps that S1, a key point detection deep neural network is trained, specifically, the key point detection deep neural network composed of three sub-networks is constructed, each sub-network comprises a plurality of 3 * 3 convolution kernels with the step length ranging from 1 to 2, and feature fusion is conducted between the sub-networks through the multi-size feature map fusion technology; S2, outputting a to-be-detected image to the training key point detection deep neural network, key points are detected by using the key point detection deep neural network obtained by training; the method is applied to a convolutional neural network structure for detecting the key points of the bearing workpart, and the network structure comprises a novel feature map fusion technology and a target loss function, so that the key point detection accuracy can be effectively improved.

A method for determining a predicted risk level of a clinical endpoint for a predetermined time period for a patient is provided by the present disclosure. The method includes receiving video frames of a heart, the video frames being associated with the patient, receiving electronic health record data including a number of variables associated with the patient, providing the video frames and the electronic health record data to the trained neural network, receiving a risk score from the trained neural network, and outputting a report based on the risk score to at least one of a display or a memory.

The invention relates to a visual-angle-independent video three-dimensional human body posture recognition method. The method comprises the steps: 1, a virtual data generation stage: generating a two-dimensional/three-dimensional data tuple after synthesizing virtual camera parameters based on any human body posture data set containing three-dimensional marks; step 2, a model training stage: utilizing the generated two-dimensional/three-dimensional data tuples to train a modular neural network first module used for obtaining a model with camera visual angle generalization capability and a modular neural network second module used for obtaining a model capable of protecting inter-frame action continuity respectively; and 3, an unconstrained video reasoning stage: for any unconstrained acquired video, predicting by utilizing the multi-module deep neural network obtained by training in the step 2 to obtain a three-dimensional human body posture recognition result. Compared with the priorart, the method is based on a modular neural network combination training method, and the generalization ability of three-dimensional human body posture recognition is effectively improved.

The invention discloses a DR image pulmonary tuberculosis intelligent segmentation and detection method based on deep learning, and the method comprises the steps: extracting a mask from an original DR image, filtering a small-area region, carrying out convex hull operation, and carrying out texpansion operation of the mask after the convex hull operation, and obtaining a lung effective region; labeling the original DR image, and labeling the position of a lesion area of the DR image with pulmonary tuberculosis expression; sending the masked image and the annotation information containing thepulmonary tuberculosis change area position into a deep neural network for training, and learning features of the pulmonary tuberculosis image; and using the trained model to predict the untrained DRimage, outputting the position of the pulmonary tuberculosis change area, and giving the probability of the pulmonary tuberculosis change area. According to the invention, the detection rate of pulmonary tuberculosis is greatly improved, the position of a pulmonary tuberculosis change area is accurately output, and reliable and efficient reference is provided for screening and treatment of pulmonary tuberculosis.

The invention discloses a steel bar quantity detection method and system based on a deep neural network. The method comprises the steps: generating a pre-training model according to a third-party object detection data set and based on an improved Faster RCNN algorithm; inputting a steel bar data set including labeling information, performing first data enhancement conversion on pictures in the steel bar data set, and performing transfer learning training on the pre-training model to generate a steel bar detection model; acquiring a to-be-detected picture including the cross section of the steel bar, obtaining a first detection frame through the steel bar detection model, performing second data enhancement conversion on the to-be-detected picture to obtain a plurality of second enhanced pictures, and obtaining a second detection frame through the steel bar detection model; matching the first detection frame and the second detection frame according to the coordinate information, and obtaining the detection number of the steel bars based on the confidence degree of the steel bars and the overlapping degree of the detection frames. The steel bar detection and warehousing speed is increased, manpower and material resources are saved, and the detection accuracy is improved.

The invention discloses a vehicle load dynamic real-time calculation method, which comprises the following steps of: firstly, acquiring vehicle power system data in real time through a sensor, and establishing a vehicle dynamic theoretical model; introducing a deep neural network model, reducing the influence of sensor noise and environmental noise by using big data, and filtering interference factors to obtain a neural network model; wherein the big data comprises full-life-cycle time series data of a plurality of automobiles; training a neural network model at the cloud through a GPU cluster, and realizing edge deployment after test optimization; and finally, acquiring sensor data in real time through a vehicle-mounted module, synchronously inputting the sensor data into the trained neural network model, and calculating dynamic vehicle load data. Real-time monitoring of commercial vehicle load data is achieved, the overload phenomenon can be effectively eradicated, good social benefits are achieved. A pure software scheme is adopted, and compared with a traditional fixed-point inspection station and a refitted vehicle scheme, the invention has overwhelming advantages.

The invention discloses a hybrid learning-based adaptive real-time energy management method for a more-electric aircraft. The method comprises the steps of S1, establishing a more-electric aircraft energy management model; s2, in an off-line stage, acquiring load configuration data of the more-electric aircraft and inputting the load configuration data into a commercial solver in batches to obtain a solving result of the model, and forming a data set; s3, establishing an integrated deep neural network model; s4, training an integrated deep neural network model based on the data set; and S5, in an online stage, inputting real-time operation scene data into the trained model to obtain an integer solution, judging the feasibility of the integer solution, solving other continuous variables of the more-electric aircraft energy management model by adopting a commercial solver when the integer solution is feasible, otherwise, abandoning the integer solution, and obtaining all variables of the more-electric aircraft energy management model by adopting the commercial solver. The method can adapt to the dynamic operation scene of the more-electric aircraft, reduces the occupation of computing resources, reduces the solving time, improves the energy scheduling rate, and has the real-time performance, the self-adaption performance and the optimality.

The invention discloses an image nonlinear interpolation acquisition method and system based on deep learning, and belongs to the technical field of image processing, and the method comprises the steps: 1, building a data set; 2, detecting an unqualified target in the interpolated image by using a deep learning target detection method to obtain a first target detection deep neural network; 3, detecting the position of an unqualified target in the original image and pixels near the unqualified target by using a deep learning target detection method, and obtaining a second target detection deepneural network; 4, based on the data set, the first target detection deep neural network and the second target detection deep neural network, performing training to obtain a deep neural network, and outputting a correct interpolation through the deep neural network. According to the invention, a deep learning target detection technology is used, a data set of artificial interpolation correction isconstructed, an image background is combined, an unqualified target and related nearby related pixels in a traditional interpolation algorithm are identified, and a correct interpolation is output incombination with an example of artificial interpolation.

The invention discloses a working parameter control method for edge computing based on the Internet of Things. The development of the Internet of Things technology and the edge technology provides a new underlying support for the smart industry. In the method, an industrial production line for dedusting an electronic component is constructed based on the Internet of Things and edge calculation, a nozzle for dedusting is connected to an edge end in the Internet of Things, and working parameter control is performed by adopting an artificial intelligence control mode based on deep learning based on the calculation performance of the edge end. Moreover, for improving the calculation speed of the edge end, training and inferring of a convolutional neural network and statistical processing of the size data of the electronic component are carried out at the cloud end of the Internet of Things, so that category data about the surface dust amount of the electronic component and category data about the size of the electronic component are obtained. Therefore, when working parameter control is carried out at the edge end, the parameter quantity of the deep neural network used by the edge end can be reduced, so that the calculation quantity of the edge end is reduced.

The invention discloses an image reconstruction method and device for capacitive coupling electrical impedance tomography, and the method comprises the steps: taking an unsupervised deep convolutional network as prior information to restrain an image iteration reconstruction process of capacitive coupling electrical impedance tomography, and training a deep neural network with random initialization network parameters, according to the method, the network can learn intrinsic hidden information of the network from a noise label image, a local optimal solution of the noise image is gradually found through operations such as alternate weighted average summation in the training process, a result similar to a real image is generated, and denoising is completed. According to the network prior applied to capacitive coupling electrical impedance tomography, a mode similar to manual prior can be formed to carry out noise constraint on image reconstruction, intelligent parameter adjustment can be carried out according to the structural characteristics of the image, and a reconstructed image with higher quality is formed.

The invention relates to the field of security center platforms. The security center platform makes risk control strategies for users, defines user risks based on historical data and real-time behavior data, encapsulates behavior indexes, encapsulates strategies, and intercepts and marks behavior results. Specifically, the invention provides a security center platform system; the system adopts a label-based classification technology using a deep neural network model to calculate weight values of different security functions in each service scene so as to integrate the security functions in different service scenes; and by means of the mode, the different schemes can be customized according to different travel scenes, the development cost is greatly reduced, and meanwhile, the customization requirement of an operator is met.

The invention discloses a live shopping interestingness prediction method based on eye movement characteristics and DeepFM. Data processing related to fixation time is carried out on a live video output by an eye tracker. And tracking the live selling goods by using the packaged tracking model. And establishing a tracking frame as a user viewpoint and a target area. On the basis of the obtained eye movement data of the user, a cooperation information graph is introduced, and user behaviors and project knowledge are coded into a unified relation graph through the cooperation information graph; based on a DeepFM architecture, adding a self-attention mechanism on a deep neural network to improve the learning ability of the model for key information; and outputting a result, and judging the precision of the model through a binary cross entropy loss function Loss and an AUC. A model is trained based on historical browsing data (including eye movement data) of a user. A user can predict the interestingness of the user for live broadcast commodities through the model, so that related personnel can adjust a live broadcast strategy, and the user experience in live broadcast is improved.

The embodiment of the invention provides an object statistical model training method, and an object statistical method and device. The training method comprises: obtaining a training sample data set, wherein training samples in the training sample data set comprise training images and label data of the training images, the training images comprise multiple objects, and the objects comprise animals and/or plants; inputting the training image into the deep neural network model, and outputting a prediction density map; calculating a loss function according to the prediction density map and the label data to obtain a loss result; and iteratively adjusting network parameters of the deep neural network model according to the loss result, and generating a trained object statistical model.

The invention discloses an information recommendation method based on deep learning. The method comprises the following steps: S1, inputting a data set into a recommendation system; and S2, in the recommendation system, performing data processing on the data set in a recall stage, a sorting stage and a feature processing stage of the deep learning model in sequence to obtain an information recommendation result. An improved probability matrix decomposition model is provided, personal interests, preferences and the like of consumers or commodities are integrated, and purposiveness and accuracy of RS recall can be improved; meanwhile, an FWHFM model based on the deep neural network is constructed, an FHWFM interaction layer generated by a feature domain pair embedding matrix and a feature embedding vector is spliced with an original feature vector layer, training is carried out through the deep neural network DNN, and the click rate estimation effect is obviously superior to that of other deep models; and based on the improvement of the two models in the recommendation system, the accuracy of information recommendation in the information recommendation process is improved.

The invention discloses an electric power spot service system fault prediction method and device, computer equipment and a storage medium, and the method comprises the steps: receiving real-time equipment parameter time sequence data of hardware equipment in an electric power spot service system, inputting a first type of deep neural network to obtain the fault probability of the hardware equipment, wherein the first-class deep neural network is obtained through pre-construction, and comprises the steps of receiving equipment parameter time sequence data of hardware equipment in an electric power spot service system before a fault, and constructing an equipment parameter dictionary according to the equipment parameter time sequence data; performing vectorization processing on the equipment parameters by using the equipment parameter dictionary to obtain vectorized equipment parameter time sequence data and training data; and carrying out fine-tuning transfer learning on the training data by using the deep neural network to obtain the first-class deep neural network. According to the invention, safe operation and reliable operation of an electric power spot market and a business system are guaranteed.