7779 results about "Deep neural networks" patented technology

A method and system for anatomical object detection using marginal space deep neural networks is disclosed. The pose parameter space for an anatomical object is divided into a series of marginal search spaces with increasing dimensionality. A respective sparse deep neural network is trained for each of the marginal search spaces, resulting in a series of trained sparse deep neural networks. Each of the trained sparse deep neural networks is trained by injecting sparsity into a deep neural network by removing filter weights of the deep neural network.

The invention discloses a vehicle type recognition method based on a rapid R-CNN deep neural network, which mainly comprises unsupervised deep learning, a multilayer CNN (Convolutional Neural Network), a regional advice network, network sharing and a softmax classifier. The vehicle type recognition method realizes a framework for implementing end-to-end vehicle detection and recognition by using one rapid R-CNN network in a real sense, and is capable of carrying out quick vehicle sub-category recognition with high accuracy and robustness under the environment of being applicable to the shape diversity, the illumination variation diversity, the background diversity and the like of vehicle targets.

Methods and systems for image registration using an intelligent artificial agent are disclosed. In an intelligent artificial agent based registration method, a current state observation of an artificial agent is determined based on the medical images to be registered and current transformation parameters. Action-values are calculated for a plurality of actions available to the artificial agent based on the current state observation using a machine learning based model, such as a trained deep neural network (DNN). The actions correspond to predetermined adjustments of the transformation parameters. An action having a highest action-value is selected from the plurality of actions and the transformation parameters are adjusted by the predetermined adjustment corresponding to the selected action. The determining, calculating, and selecting steps are repeated for a plurality of iterations, and the medical images are registered using final transformation parameters resulting from the plurality of iterations.

A method and system for determining fractional flow reserve (FFR) for a coronary artery stenosis of a patient is disclosed. In one embodiment, medical image data of the patient including the stenosis is received, a set of features for the stenosis is extracted from the medical image data of the patient, and an FFR value for the stenosis is determined based on the extracted set of features using a trained machine-learning based mapping. In another embodiment, a medical image of the patient including the stenosis of interest is received, image patches corresponding to the stenosis of interest and a coronary tree of the patient are detected, an FFR value for the stenosis of interest is determined using a trained deep neural network regressor applied directly to the detected image patches.

A method and system for anatomical object detection using marginal space deep neural networks is disclosed. The pose parameter space for an anatomical object is divided into a series of marginal search spaces with increasing dimensionality. A respective deep neural network is trained for each of the marginal search spaces, resulting in a series of trained deep neural networks. Each of the trained deep neural networks can evaluate hypotheses in a current parameter space using discriminative classification or a regression function. An anatomical object is detected in a medical image by sequentially applying the series of trained deep neural networks to the medical image.

The invention provides a calculation apparatus and method for an accelerator chip accelerating a deep neural network algorithm. The apparatus comprises a vector addition processor module, a vector function value calculator module and a vector multiplier-adder module, wherein the vector addition processor module performs vector addition or subtraction and / or vectorized operation of a pooling layer algorithm in the deep neural network algorithm; the vector function value calculator module performs vectorized operation of a nonlinear value in the deep neural network algorithm; the vector multiplier-adder module performs vector multiplication and addition operations; the three modules execute programmable instructions and interact to calculate a neuron value and a network output result of a neural network and a synaptic weight variation representing the effect intensity of input layer neurons to output layer neurons; and an intermediate value storage region is arranged in each of the three modules and a main memory is subjected to reading and writing operations. Therefore, the intermediate value reading and writing frequencies of the main memory can be reduced, the energy consumption of the accelerator chip can be reduced, and the problems of data missing and replacement in a data processing process can be avoided.

The present invention is directed to systems and methods for detecting objects in a radar image stream. Embodiments of the invention can receive a data stream from radar sensors and use a deep neural network to convert the received data stream into a set of semantic labels, where each semantic label corresponds to an object in the radar data stream that the deep neural network has identified. Processing units running the deep neural network may be collocated onboard an airborne vehicle along with the radar sensor(s). The processing units can be configured with powerful, high-speed graphics processing units or field-programmable gate arrays that are low in size, weight, and power requirements. Embodiments of the invention are also directed to providing innovative advances to object recognition training systems that utilize a detector and an object recognition cascade to analyze radar image streams in real time. The object recognition cascade can comprise at least one recognizer that receives a non-background stream of image patches from a detector and automatically assigns one or more semantic labels to each non-background image patch. In some embodiments, a separate recognizer for the background analysis of patches may also be incorporated. There may be multiple detectors and multiple recognizers, depending on the design of the cascade. Embodiments of the invention also include novel methods to tailor deep neural network algorithms to successfully process radar imagery, utilizing techniques such as normalization, sampling, data augmentation, foveation, cascade architectures, and label harmonization.

The technology disclosed proposes using a combination of computationally cheap, less-accurate bag of words (BoW) model and computationally expensive, more-accurate long short-term memory (LSTM) model to perform natural processing tasks such as sentiment analysis. The use of cheap, less-accurate BoW model is referred to herein as “skimming”. The use of expensive, more-accurate LSTM model is referred to herein as “reading”. The technology disclosed presents a probability-based guider (PBG). PBG combines the use of BoW model and the LSTM model. PBG uses a probability thresholding strategy to determine, based on the results of the BoW model, whether to invoke the LSTM model for reliably classifying a sentence as positive or negative. The technology disclosed also presents a deep neural network-based decision network (DDN) that is trained to learn the relationship between the BoW model and the LSTM model and to invoke only one of the two models.

A method and system for approximating a deep neural network for anatomical object detection is discloses. A deep neural network is trained to detect an anatomical object in medical images. An approximation of the trained deep neural network is calculated that reduces the computational complexity of the trained deep neural network. The anatomical object is detected in an input medical image of a patient using the approximation of the trained deep neural network.

Face recognition of a face, to determine whether the face correlates with an enrolled face, may include generating a personalized three-dimensional (3D) face model based on a two-dimensional (2D) input image of the face, acquiring 3D shape information and a normalized 2D input image of the face based on the personalized 3D face model, generating feature information based on the 3D shape information and pixel color values of the normalized 2D input image, and comparing the feature information with feature information associated with the enrolled face. The feature information may include first and second feature information generated based on applying first and second deep neural network models to the pixel color values of the normalized 2D input image and the 3D shape information, respectively. The personalized 3D face model may be generated based on transforming a generic 3D face model based on landmarks detected in the 2D input image.

The invention provides a multi-mode based emotion recognition method, comprising a data collection device, an output device and an emotion analysis software system, wherein the emotion analysis software system performs comprehensive analysis and reasoning on data obtained by the data collection device and finally outputs a result to the output device. The specific steps are as follows: an emotionrecognition step based on facial image expressions, an emotion recognition step based on voice signals, an emotion analysis step based on text semantic, an emotion recognition step based on human gestures, an emotion recognition step based on physiological signals, a semantic comprehension step based on multi-round dialogues, and a multi-mode emotion semantic fusion association judgment step basedon timing sequence. The multi-mode based emotion recognition method provided by the invention has the advantages of breaking through the five kinds of single-mode emotion recognition, innovatively performing comprehensive judgment on the information of multiple single modes by using a deep neural network through neural network coding, deep correlation and understanding, greatly improving the accuracy and being suitable for most general inquiry interaction application scenes.

The embodiment of the invention discloses a lane line data processing method and a lane line data processing device. The method comprises the following steps of: obtaining an original image and positioning data of the original image; using a deep neural network model for calculating the pixel confidence degree of each pixel, conforming to lane line features, in the original image; determining the lane line outlines from the original image, and using the lane line outlines as candidate lane line; calculating the lane line confidence degree of the candidate lane lines; screening the candidate lane lines according to the lane line confidence degree of the candidate lane lines; recognizing the attribute information of the lane line by aiming at the screened lane line; and determining map data of the lane line according to the attribute information of the lane line and positioning data during the original image shooting. The lane line data processing method and the lane line data processing device provided by the embodiment have the advantages that the lane line data can be efficiently and precisely determined; the labor cost in high-precision map production is greatly reduced; and the large-scale high-precision map production can be realized.

The technology disclosed provides a so-called “joint many-task neural network model” to solve a variety of increasingly complex natural language processing (NLP) tasks using growing depth of layers in a single end-to-end model. The model is successively trained by considering linguistic hierarchies, directly connecting word representations to all model layers, explicitly using predictions in lower tasks, and applying a so-called “successive regularization” technique to prevent catastrophic forgetting. Three examples of lower level model layers are part-of-speech (POS) tagging layer, chunking layer, and dependency parsing layer. Two examples of higher level model layers are semantic relatedness layer and textual entailment layer. The model achieves the state-of-the-art results on chunking, dependency parsing, semantic relatedness and textual entailment.

The invention relates to a deep learning-based question and answer matching method. The method comprises the following steps of: 1) sufficiently learning word orders and sentence local features of a question text and an answer text by utilizing two underlying deep neural networks: a long short-term memory network LSTM and a convolutional neural network CNN; and 2) selecting a keyword with best semantic matching on the basis of a pooling manner of an attention mechanism AM. Compared with existing methods, the method has the advantages of being in low in feature engineering workload, strong in cross-field performance and relatively high in correctness, and can be effectively applied to the fields of commercial intelligent customer service robots, automatic driving, internet medical treatment, online forum and community question answering.

The invention discloses a method and system for deep learning algorithm acceleration on a field-programmable gate array platform. The field-programmable gate array platform is composed of a universal processor, a field-programmable gate array and a storage module. The method comprises: according to a deep learning prediction process and a training process, a general computation part that can be operated on a field-programmable gate array platform is determined by combining a deep neural network and a convolutional neural network; a software and hardware cooperative computing way is determined based on the determined general computation part; and according to computing logic resources and the bandwidth situation of the FPGA, the number and type of IP core solidification are determined, and acceleration is carried out on the field-programmable gate array platform by using a hardware computing unit. Therefore, a hardware processing unit for deep learning algorithm acceleration is designed rapidly based on hardware resources; and compared with the general processor, the processing unit has characteristics of excellent performance and low power consumption.

The invention discloses a multi-scale image semantic segmentation method. The method comprises the following steps: obtaining a to-be-segmented image and a corresponding label; constructing a full convolutional deep neural network, wherein the full convolutional deep neural network comprises a convolution module, a hole convolution module, a pyramid pooling module, a 1 * 1 * depth convolution layer and a deconvolution structure; setting hole convolution as channel-by-channel operation, and utilizing low-scale, medium-scale and high-scale characteristics in a targeted mode; training the full convolutional deep neural network, establishing a loss function, and determining parameters of the full convolutional deep neural network by training the sample image; and inputting the to-be-segmentedimage into the trained full convolutional deep neural network to obtain a semantic segmentation result. By means of the method, the image semantic segmentation problem with complex details, holes andlarge targets can be well solved while the calculated amount and the parameter number are reduced, and the consistency of category labels can be reserved while the target edges can be well segmented.

An artificial neural network for analyzing input data, the input data being a 3D tensor having D channels, such as D frames of a video snippet, to recognize an action therein, including: D spatial transformer modules, each generating first and second spatial transformations and corresponding first and second attention windows using only one of the D channels, and transforming first and second regions of each of the D channels corresponding to the first and second attention windows to generate first and second patch sequences; first and second CNNs, respectively processing a concatenation of the D first patch sequences and a concatenation of the D second patch sequences; and a classification network receiving a concatenation of the outputs of the first and second CNNs and the D sets of transformation parameters of the first transformation outputted by the D spatial transformer modules, to generate a predicted action class.

The invention relates to the field of image recognition, and especially relates to a complex character recognition method based on deep learning. Through the analysis of character complexity, a training sample, which contains a to-be-recognized image noise model and a distortion characteristic model, generated by a random sample generator is employed for the training of a deep neural network. The training sample comprises complex noise and distortion, and can meet the demands of the recognition of various types of complex characters. A few of manually annotated first training sample sets and a large amount of randomly generated second training sample sets are mixed and then inputted to the deep neural network, thereby solving a problem that a large number of manually annotated training samples are needed for character recognition through the deep neural network. Moreover, the most advanced deep neural network is employed for automatic learning under the condition that the noise and distortion of a to-be-recognized image are retained, thereby avoiding information loss caused by noise reduction in a conventional OCR method, and improving the recognition accuracy.

Systems and methods for detecting patterns in data from a time-series are provided. In one implementation, a method for pattern detection includes obtaining data in a time-series and creating one-dimensional or multi-dimensional windows from the time-series data. The one-dimensional or multi-dimensional windows are created either independently or jointly with the time-series. The method also includes training a deep neural network with the one-dimensional or multi-dimensional windows utilizing historical and / or simulated data to provide a neural network model. Also, the method includes processing ongoing data with the neural network model to detect one or more patterns of a particular category in the ongoing data, and localizing the one or more patterns in time.

A method for predicting a canonical form for an input text sequence includes predicting the canonical form with a neural network model. The model includes an encoder, which generates a first representation of the input text sequence based on a representation of n-grams in the text sequence and a second representation of the input text sequence generated by a first neural network. The model also includes a decoder which sequentially predicts terms of the canonical form based on the first and second representations and a predicted prefix of the canonical form. The canonical form can be used, for example, to query a knowledge base or to generate a next utterance in a discourse.

The invention discloses a mobile ad hoc network intrusion detection method and device based on deep learning, relating to the field of wireless network safety. The device comprises a data acquisition module, a data fusion module, a preprocessing module, a storage module, an intrusion detection module and a response warning module. After fusion and redundancy elimination of captured wireless data packages, network behavior features are extracted and stored; after deep learning of the network behavior features, a deep neural network model expressing network behaviors is established; and to-be-detected network data is input into the deep neural network model, after intrusion is judged and recognized, response and warning are performed. According to the method, network behavior feature vectors which are detected and are considered to be abnormal are stored and are used for training the deep neutral network. When occurring again, the intrusion types can be detected and recognized. While the model training and detection efficiency are guaranteed, the detection accuracy is improved, and the safety of the mobile ad hoc network is further improved.

The technology disclosed proposes using a combination of computationally cheap, less-accurate bag of words (BoW) model and computationally expensive, more-accurate long short-term memory (LSTM) model to perform natural processing tasks such as sentiment analysis. The use of cheap, less-accurate BoW model is referred to herein as “skimming”. The use of expensive, more-accurate LSTM model is referred to herein as “reading”. The technology disclosed presents a probability-based guider (PBG). PBG combines the use of BoW model and the LSTM model. PBG uses a probability thresholding strategy to determine, based on the results of the BoW model, whether to invoke the LSTM model for reliably classifying a sentence as positive or negative. The technology disclosed also presents a deep neural network-based decision network (DDN) that is trained to learn the relationship between the BoW model and the LSTM model and to invoke only one of the two models.

The invention discloses a named entities recognition method based on bidirectional LSTM and CRF. The named entities recognition method based on the bidirectional LSTM and CRF is improved and optimizedbased on the traditional named entities recognition algorithm in the prior art. The named entities recognition method based on the bidirectional LSTM and CRF comprises the following steps: (1) preprocessing a text, extracting phrase information and character information of the text; (2) coding the text character information by means of the bidirectional LSTM neural network to convert the text character information into character vectors; (3) using the glove model to code the text phrase information into word vectors; (4) combining the character vectors and the word vectors into a context information vector and putting the context information vector into the bidirectional LSTM neural network; and (5) decoding the output of the bidirectional LSTM with a linear chain condition random field to obtain a text annotation entity. The invention uses a deep neural network to extract text features and decodes the textual features with the condition random field, therefore, the text feature information can be effectively extracted and good effects can be achieved in the entity recognition tasks of different languages.

A system is provided for unsupervised cross-domain image generation relative to a first and second image domain that each include real images. A first generator generates synthetic images similar to real images in the second domain while including a semantic content of real images in the first domain. A second generator generates synthetic images similar to real images in the first domain while including a semantic content of real images in the second domain. A first discriminator discriminates real images in the first domain against synthetic images generated by the second generator. A second discriminator discriminates real images in the second domain against synthetic images generated by the first generator. The discriminators and generators are deep neural networks and respectively form a generative network and a discriminative network in a cyclic GAN framework configured to increase an error rate of the discriminative network to improve synthetic image quality.

The invention discloses a deep neural network and reinforcement learning-based generative machine reading comprehension method. According to the method, texts and questions are encoded through an attention mechanism-combined deep neural network so as to form question information-fused text vector expressions, and decoding is carried out through a unidirectional LSTM decoder so as to gradually generate corresponding answer texts. According to the reading comprehension method, the advantages of extractive models and generative models are fused, training is carried out by adoption of a multi-taskcombined optimization manner, and a reinforcement learning method is used in the training process, so that benefit is brought to generate more correct and fluent answer texts.

The invention discloses a power load forecasting method based on a long short term memory (LSTM) neuron network. The power load forecasting method comprises inputting power load data and a region feature factor at a historic moment through an input unit; carrying out training and modeling on the power load data and the region feature factor at the historic moment by means of an LSTM network in order to generate a deep neural network load forecasting model which is a single-layer multi-task deep neural network model or a double-layer multi-task deep neural network model used for power supply load forecasting; forecasting the power load in an area needing to be forecasted by means of the deep neural network load forecasting model, and generating a forecasting result of the power load in the area; and outputting the forecasting result of the power load in the area through an output unit. According to the invention, a multi-task learning power load forecasting model is constructed based on the LSTM network in the deep learning field, power consumption loads in multiple areas can be forecasted accurately, and the forecasting effect is improved.

The present invention relates to a method and an apparatus for processing lane line data, a computer device and a storage medium. The method includes: acquiring and dividing three-dimensional point cloud data of a to-be-processed road; and processing the three-dimensional point cloud data of each segment after the segmentation to obtain a two-dimensional gray image of the three-dimensional point cloud data of each segment; using a pre-trained deep neural network model, respectively extracting a lane line region and a lane line attributed such as a dotted or full lane line in each two-dimensional gray image to obtain a corresponding lane line region map; and according to the three-dimensional point cloud data corresponding to the lane line region map and the lane line attribute such as thedotted or full lane line, splicing the lane line region map to obtain lane line data of the to-be-processed road. By adopting this method, the processing efficiency is improved, and compared with ordinary machine learning, it is not easily affected by interference items such as characters and cars in the three-dimensional point cloud data, and the accuracy of lane line region extraction is improved.

The present invention discloses a generative adversarial network-based multi-pose face generation method. According to the generative adversarial network-based multi-pose face generation method, in a training phase, the face data of various poses are collected; two deep neural networks G and D are trained on the basis of a generative adversarial network; and after training is completed, the generative network G is inputted on the basis of random sampling and pose control parameters, so that face images of various poses can be obtained. With the method of the invention adopted, a large quantity of different face images of a plurality of poses can be generated, and the problem of data shortage in the multi-pose face recognition field can be solved; the newly generated face images of various poses are adopted as training data to train an encoder for extracting the identity information of the images; in a final testing process, an image of a random pose is inputted, and identity information features are obtained through the trained encoder; and the face images of various poses of the same person are obtained through the trained generative network.

The present invention discloses a deep learning method-based short-term traffic flow prediction method. The influence of the traffic flow rate change of the neighbor points of a prediction point, the time characteristic of the prediction point and the influence of the periodic characteristic of the prediction point on the traffic flow rate of the prediction point are considered simultaneously. According to the deep learning method-based short-term traffic flow prediction method of the invention, a convolutional neural network and a long and short-term memory (LSTM) recurrent neural network are combined to construct a Conv-LSTM deep neural network model; a two-way LSTM model is used to analyze the traffic flow historical data of the point and extract the periodic characteristic of the point; and a traffic flow trend and a periodic characteristic which are obtained through analysis are fused, so that the prediction of traffic flow can be realized. With the method of the invention adopted, the defect of the incapability of an existing method to make full use of time and space characteristics can be eliminated, the time and space characteristics of the traffic flow are fully extracted, and the periodic characteristic of the data of the traffic flow is fused with the time and space characteristics, and therefore, the accuracy of short-term traffic flow prediction results can be improved.

A system and method for using a data genome to identify suspicious financial transactions. In one embodiment, the method comprises receiving a data set of financial activity data of multiple participants; configuring a deep neural network and thresholds, wherein the thresholds enable detection of what is within abnormal range of financial activity, patterns, and behavior over a period of time; converting the data set to a genome containing a node for each participant among the multiple participants; computing threat vectors for each node within a graphical representation of the genome that represents behavioral patterns of participants in financial activities, including determining when a key risk indicator (KRI) value computed for a particular threshold within the data set falls outside of a dynamically determined range bounded by thresholds, wherein the threat vectors automatically identify one or more of suspicious participants and suspicious activities in a provided financial activity pattern; and determining a particular edge in the network whose behavior falls outside the dynamically determined range associated with normal activity as a suspicious.