587 results about "Forward propagation" patented technology

An eyepiece includes an eyepiece frame, an in-coupling polarization beam splitter (“PBS”), an end reflector, and an out-coupling PBS. The eyepiece frame defines an air cavity and includes an illumination region for receiving computer generated image (“CGI”) light into the eyepiece frame and a viewing region to be aligned with an eye of a user. The in-coupling PBS is supported within the eyepiece frame at the illumination region to re-direct the CGI light to a forward propagation path extending along the air cavity towards the viewing region. The end reflector is disposed to reflect the CGI light back along a reverse propagation path within the eyepiece frame. The out-coupling PBS is supported at the viewing region to pass the CGI light traveling along the forward propagation path and to redirect the CGI light traveling along the reverse propagation path out of an eye-ward side of the eyepiece frame.

Acoustic imaging provides images with reduced reverberation noise and images of nonlinear scattering and propagation parameters of the object. The received signal from transmitted dual frequency band acoustic pulse complexes is processed with overlapping high and low frequency pulses. The high frequency pulse is used for image reconstruction and the low frequency pulse manipulate the nonlinear scattering and / or propagation properties of the high frequency pulse. The scattered signal from a single dual band pulse complex for filtering in the fast time (depth time) domain provides a signal with suppression of reverberation noise with 1st harmonic sensitivity and increased spatial resolution. Through filtering in the pulse number coordinate and corrections for nonlinear propagation delays, a linear back scattering signal with suppressed pulse reverberation noise, a nonlinear back scattering signal, and quantitative nonlinear forward propagation and scattering parameters are extracted.

A resource delivery network and method for distributing content in the network is disclosed herein. The network comprises a plurality of servers arranged in tiers and partitioned. Each server includes a resource store with a set of resources for distribution to a successive tier. Updates to each successive tier are provided by a pull-forward client on servers in the tier. This forward propagation mechanism maximizes resource availability at edge servers in the network. Resources transmitted to the edge tier servers may be transformed, combined, and rendered without taxing lower tier servers. Transformation and pre-rendering of data can be performed by low priority CPU tasks at each layer of the system.

An eyepiece for a head mounted display includes an illumination module, an end reflector, a viewing region, and a polarization rotator. The illumination module provides CGI light along a forward propagation path within the eyepiece. The end reflector is disposed at an opposite end of the eyepiece from the illumination module to reflect the CGI light back along a reverse propagation path within the eyepiece. The viewing is disposed between the illumination module and the end reflector and includes an out-coupling polarizing beam splitter (“PBS”). The out-coupling PBS passes the CGI light traveling along the forward propagation path and redirects the CGI light traveling along the reverse propagation path out of an eye-ward side of the eyepiece. The polarization rotator is disposed in the forward and reverse propagation paths between the out-coupling PBS and the end reflector.

The invention discloses a face recognition method based deep learning and a face recognition device thereof and electronic equipment. The method comprises the steps that a convolutional neural network model is constructed, and the convolutional neural network model comprises a first convolution unit, a first pooling layer, multiple convolution combinations, a second pooling layer and full connection layers which are connected in turn, wherein the first convolution unit comprises a first convolution layer, a batch normalization layer and an excitation function layer, the excitation function layer simultaneously uses a ReLU function and a NReLU function to act as the excitation function, and the adjacent convolution combinations are connected by the short circuit layer of the residual network; and the convolutional neural network model is trained, the training data are inputted to the convolutional neural network model and training is performed by using the stochastic gradient descent method, and the last full connection layer is removed out of the trained convolutional neural network model and then only forward propagation is performed so as to act as the face feature data required for face recognition. ReLU + NReLU are used as the excitation function so that the computational burden can be reduced, the accuracy can be guaranteed, the model size can be reduced and the operation speed can be enhanced.

The invention relates to an image classification method capable of effectively preventing a convolutional neural network from being overfit. The image classification method comprises the following steps: obtaining an image training set and an image test set; training a convolutional neural network model; and carrying out image classification to the image test set by adopting the trained convolutional neural network model. The step of training the convolutional neural network model comprises the following steps: carrying out pretreatment and sample amplification to image data in the image training set to form a training sample; carrying out forward propagation to the training sample to extract image features; calculating the classification probability of each sample in a Softmax classifier; according to the probability yi, calculating to obtain a training error; successively carrying out forward counterpropagation from the last layer of the convolutional neural network by the training error; and meanwhile, revising a network weight matrix W by SGD (Stochastic Gradient Descent). Compared with the prior art, the invention has the advantages of being high in classification precision, high in rate of convergence and high in calculation efficiency.

The invention discloses an image Hash indexing and establishing method based on deep learning and belongs to the technical field of image retrieval. The method comprises the steps that an image data set is firstly divided to obtain a test sample set, a training sample set and an image library, then the characteristic that depth characteristics extracted by a deep learning model have the very good expression capability to image semantic is fully utilized to establish different depth Hash network models of two deep convolution network structures, then each image in the test sample set and the image library undergoes forward propagations of the two models respectively, corresponding two groups of initial Hash codes are calculated, then two groups of initial Hash codes of the same image are spliced and fused to serve as fused Hash codes of the image, and similar image retrieval results are obtained by calculating and querying Hamming distances of the fused Hash codes of images and each image in the image library and arranging the Hamming distances from small to large. The image Hash indexing and establishing method makes large-scale image retrieval accurate and effective.

The invention discloses a visual detection method of shockproof hammer defect detection. The visual detection method comprises the steps that denoising and anti-shaking preprocessing is performed on an aerial photographing image so as to obtain an original image to be detected; the existing original image is expanded by using the method of geometric transformation, scale change and contrast transformation so as to generate more data similar to the original image; samples are acquired, a shockproof hammer in the aerial photographing image is acquired and the size of the shockproof hammer is mainly acquired; a network model to be trained is determined, and the sample data are inputted to perform forward propagation and reverse propagation to adjust the weight so as to obtain the optimized detection network model parameters; the image to be detected is identified by using the trained model and the position of the hammer of the shockproof hammer is determined; and the lead in which the hammer is located is determined, and shockproof hammer defect discrimination is performed according to the relative position of the lead and the shockproof hammer and the constraints of respective directions.

The present invention discloses a deep learning-based insulator identification method. The insulator identification method comprises the steps of pre-processing an aerial image, and secondly, extending the data via the methods, such as the geometric transformation, the contrast enhancement, an analog noise adding method, etc.; acquiring the insulator samples, aiming at the insulators of different types, classifying to acquire; determining a to-be-trained model structure; inputting the samples in the to-be-trained model, and continuously adjusting the weights and the bias parameters by the forward propagation and backward propagation methods, and finally determining an optimal model parameter, based on the trained model, taking a to-be-detected image as an input signal, and by the network multi-layer convolution, pooling and full-connection operations, obtaining a final detection identification result. According to the present invention, by a deep learning method, the insulator characteristics are learned continuously, a learning network model is determined, the different insulators are identified under different background environments, and support is provided for the electric power maintenance decisions.

A method includes determining object class probabilities of pixels in a first input image by examining the first input image in a forward propagation direction through layers of artificial neurons of an artificial neural network. The object class probabilities indicate likelihoods that the pixels represent different types of objects in the first input image. The method also includes selecting, for each of two or more of the pixels, an object class represented by the pixel by comparing the object class probabilities of the pixels with each other, determining an error associated with the object class that is selected for each pixel of the two or more pixels, determining one or more image perturbations by back-propagating the errors associated with the object classes selected for the pixels of the first input image through the layers of the neural network without modifying the neural network, and modifying a second input image by applying the one or more image perturbations to one or more of the first input image or the second input image prior to providing the second input image to the neural network for examination by the neurons in the neural network for automated object recognition in the second input image.

The invention discloses a small target detection and recognition method for enhanced feature learning, belongs to the field of image processing, pattern recognition and computer vision, and solves theproblems of low small target detection and recognition task detection precision and low network efficiency in the prior art. The method comprises: sequentially constructing a basic network module, afeature extraction module, a candidate box generation module and a prediction output module to serve as a small target detection and identification network; preprocessing the extracted small target sample image data based on the extracted small target sample image data; inputting the preprocessed small target sample image data into the small target detection and recognition network with initialized parameters for training to obtain a trained small target detection and recognition network; and inputting a to-be-predicted small target image into the trained small target detection and recognitionnetwork, and outputting the prediction box position and category information of the small target end to end through forward propagation. The method is used for small target detection and recognition.

The invention discloses a multi-target tracking method based on a recurrent neural network, comprising the following steps: building a monitoring video data set marked with each frame of pedestrian location; manually expanding the monitoring video data set marked with each frame of pedestrian location to get training set samples; grouping the training set samples to get multiple training groups; building a multi-target tracking network; inputting the training groups in sequences to the multi-target tracking network for training; and inputting video data to be tested to the trained multi-target tracking network and performing forward propagation to get the moving trajectories of multiple targets. According to the invention, a proposed network model is trained in an end-to-end manner using raw data and a large amount of data obtained through artificial expansion, complex tasks such as data association and track estimation are completed under a unified neural network architecture, and the moving trajectories of targets can be tracked effectively under different directions, light conditions, deformations and other complex environments.

The invention discloses a radar echo extrapolation method based on a dynamic convolution neural network. The method comprises a step of offline convolutional neural network training which comprises the steps of carrying out data preprocessing on a given training image set to obtain a training sample set, initializing a dynamic convolution neural network model, training a dynamic convolution neural network by using the training sample set, calculating an output value through network forward propagation, and updating network parameters through backward propagation such that the dynamic convolution neural network converges. The method also comprises a step of online radar echo extrapolation which comprises the steps of converting a test image set into a test sample set through data preprocessing, testing the trained dynamic convolution neural network by using the test sample set, and carrying out convolution of a laster radar echo image inputted into an image sequence and a probability vector obtained in the network forward propagation to obtain a predicted radar echo extrapolation image.

New methods of ultrasound imaging are presented that provide images with reduced reverberation noise and images of nonlinear scattering and propagation parameters of the object, and estimation of corrections for wave front aberrations produced by spatial variations in the ultrasound propagation velocity. The methods are based on processing of the received signal from transmitted dual frequency band ultrasound pulse complexes with overlapping high and low frequency pulses. The high frequency pulse is used for the image reconstruction and the low frequency pulse is used to manipulate the nonlinear scattering and / or propagation properties of the high frequency pulse. A 1st method uses the scattered signal from a single dual band pulse complex for filtering in the fast time (depth time) to provide a signal with suppression of reverberation noise and with 1st harmonic sensitivity and increased spatial resolution. In other methods two or more dual band pulse complexes are transmitted where the frequency and / or the phase and / or the amplitude of the low frequency pulse vary for each transmitted pulse complex. Through filtering in the pulse number coordinate and corrections of nonlinear propagation delays and optionally also amplitudes, a linear back scattering signal with suppressed pulse reverberation noise, a nonlinear back scattering signal, and quantitative nonlinear scattering and forward propagation parameters are extracted. The reverberation suppressed signals are further useful for estimation of corrections of wave front aberrations, and especially useful with broad transmit beams for multiple parallel receive beams. Approximate estimates of aberration corrections are given. The nonlinear signal is useful for imaging of differences in tissue properties, such as micro-calcifications, in-growth of fibrous tissue or foam cells, or micro gas bubbles as found with decompression or injected as ultrasound contrast agent.

The invention discloses a deep learning based vehicle license plate position insensitive vehicle license plate recognition method, which is characterized in that a sample set with seven vehicle license plate character labels is built, and training is carried out on a deep convolution neural network. The vehicle license plate recognition method comprises the steps of carrying out preprocessing on a vehicle license plate image to be detected, converting the preprocessed vehicle license plate image into an image whose size is the same as that of a sample in the sample set, and inputting the converted image into the trained deep convolution neutral network; carrying out one time of forward propagation in the deep convolution neural network, and outputting seven labels; and acquiring Chinese characters and characters corresponding to the labels through looking up a table so as to acquire seven vehicle license plate characters. According to the invention, vehicle license plate position information can be effectively recognized through creatively sharing a convolution layer, carrying out recognition on the seven vehicle license plate characters and carrying out sample processing in a targeted manner at the same time, thereby improving the recognition rate under a complex environment, and finally achieving a good global recognition effect.

The invention relates to an image semantic segmentation method based on a multi-layer information fusion full convolutional neural network. The method comprises the following steps: constructing the multi-information fusion full convolutional neural network; Preprocessing the training picture of the data set, and inputting a preprocessing result into a neural network for training to obtain an image semantic segmentation model; Inputting the to-be-tested image into the image semantic segmentation model, performing one-time forward propagation, and outputting a semantic segmentation result in anend-to-end manner. Basic depth features of the image are extracted by using a convolutional neural network model; According to the method, the features are divided into the low-level features and thehigh-level features, and the low-level features and the high-level features are fused into the enhanced depth features, so that the problems of low image semantic segmentation accuracy and image spatial detail information loss can be effectively solved, and finally, a better image semantic segmentation result is obtained.

The invention relates to a differential-training-based voice wake-up customization optimization method and system. After acoustic data with marks are given, neural network forward propagation is carried out by combining an initial neural-network-based acoustic model; a modeling probability for wake-up word marking is obtained by combining a marking sequence and a frame-by-frame acoustic probability; statistics is carried out on lots of texts to obtain a phoneme-level language model and a corresponding searching network is constructed; wake-up word competition item modeling is carried out based on the searching space of the phoneme-level language model and a frame-by-frame acoustic model, thereby obtaining a posterior probability; differential training of the acoustic model is carried out by combining the wake-up word competition item modeling and the wake-up word marking modeling; and the acoustic model obtained by the differential training determines whether to carry out a wake-up operation by comparison of a wake-up word confidence coefficient with an automatic estimation threshold value.

A light guide includes a light guide layer having a transversely oriented side-end surface that forms a primary output aperture (exit) for light traveling in a forward propagation direction out of the end surface of the light guide (for, e.g., CPV applications) and, which forms a primary input aperture (entrance) for light traveling in a rearward propagation direction into the end surface of the light guide (for, e.g., illuminator applications), and a first plurality of light injection elements stepped (staggered) in a forward light propagation direction in a first plane along lines parallel to the side-end surface or clocked (tilted) about a y-axis in a z-axis-light propagation direction in a respective first plane, wherein the light injection elements are disposed along parallel lines normal to the side-end surface. The light guide component may further comprise at least a second plurality of light injection elements stepped in at least second plane. A light guide system includes a component light guide, a lenslet array disposed adjacent a top surface of the light guide, and a light-transmitting, TIR medium layer disposed immediately adjacent at least one of the top and bottom surfaces of the light guide.

The invention discloses a traffic jam judgment method based on deep learning. The traffic jam judgment method comprises the following steps: 1, acquiring a training sample, and adding a tag so as to obtain an image which comprises the tag and corresponds to a monitoring video file; 2, performing forward propagation, namely, transmitting the image which comprises the tag and corresponds to the monitoring video file into a designed convolution neural network model, and performing forward propagation so as to obtain a type tag output by the convolution neural network model; 3, performing back propagation, namely, calculating a loss function value of the type tag output in forward propagation and an actual type tag of the sample, performing back propagation on the loss function value in a minimized error direction so as to adjust a weight matrix of a convolution layer and obtain a final convolution neural network model; 4, judging traffic jam, namely, transmitting at least one frame of image corresponding to a current monitoring video file of a selected road section into the trained final convolution neural network model, and performing forward propagation. By adopting the traffic jam judgment method, the traffic jam grade can be provided according to the traffic situation of a current road, and relatively good applicability and robustness can be achieved.

The invention relates to a Fourier parallel magnetic resonance imaging method based on a one-dimensional part of a deep convolutional network, and belongs to the technical field of magnetic resonance imaging. The method comprises the following steps: a sample set for training and a sample label set are created; an initial deep convolutional network is built; a training sample of the sample set is input into an initial deep convolutional network model to perform forward propagation, an output result of the forward propagation is compared with an expect result in the sample label set, and training is performed using a gradient descent algorithm until various layer parameters maximizing the consistency between the output result and the expect result are obtained; an optimal deep convolutional network model is established by utilizing the obtained various layer parameters; and a multi-coil under-sampling image obtained through online sampling is input into the optimal deep convolutional network model, forward propagation is performed on the optimal deep convolutional network model, and a rebuilt single-channel whole-sampling image is output. A noise of the rebuilt image can be removed well, a magnetic resonance image having a good visual effect is rebuilt, and the Fourier parallel magnetic resonance imaging method has high practical value.

The invention belongs to the technical field of medical image intelligent processing, in particular to a gastrointestinal endoscopic image classification and early cancer detection system based on a multi-task neural network. The system of the invention comprises: (1) a feature extraction backbone network; (2) Classification of gastrointestinal endoscopic images; (3) regional detection branch of early gastrointestinal cancer. The invention adopts a multi-task depth neural network structure, and classifies and detects two tasks to share a plurality of convolution layers. Endoscopic images are inputted into the neural network model, and the detection and classification results can be obtained simultaneously after a forward propagation, which can effectively reduce the computational load andimprove the classification and detection accuracy. The experimental results show that the invention can accurately classify the endoscopic images into normal and early cancers and detect the irregularlesion areas in the early cancers images, reduce the influence of human factors and improve the efficiency of clinical diagnosis.

The invention discloses a clothes classifying method based on a convolutional neural network. The method comprises the following steps of acquiring clothes image samples, and dividing the samples into training samples and testing samples; preprocessing the training samples and the testing samples; constructing a convolutional neural network model; performing training of two stages including a forward propagation stage and a backward propagation stage on the convolutional neural network model through preprocessed training samples, finishing the training when the error calculated during the backward propagation stage reaches a desired value, and acquiring a parameter of the convolutional neural network model; testing the preprocessed testing samples by using the trained convolutional neural network model and outputting final clothes classifying results. The convolutional neural network model can make clothes images directly serve as network inputs, extract image features in an implicit way and establish a global feature expression. Compared with a manually designed feature extraction way, the method is more convenient and accurate. The problem that a conventional algorithm leads to low clothes classifying accuracy is solved.

The invention discloses an image semantic segmentation method based on a depth neural network, comprising the following steps: S1, constructing a depth neural network based on a dense connection network and a pooling of a porous space pyramid; S2, pre-processing the training picture of the data set, and inputting the pre-processing result into the depth neural network for training to obtain the image semantic segmentation model; S3, inputting the image to be tested into the image semantic segmentation model, performing forward propagation, and outputting the semantic segmentation result end-to-end. The densely connected network does not desample the image, but keeps the spatial dimension of the image for training, and the multi-scale information of the image is captured by the multi-aperture spatial pyramid pooling module, which can effectively solve the problem of low accuracy of image semantic segmentation and loss of spatial detail information of the image, and finally obtain a better image semantic segmentation result.

A wave transmission medium includes an input port 3-1 and an output port 3-2. A field distribution 1 and a field distribution 2 are obtained by numerical calculations. The field distribution 1 is a field distribution of the propagation light (forward propagation light) launched into the input port 3-1. The field distribution 2 is a field distribution of the phase conjugate light (reverse propagation light) resulting from reversely transmitting from the output port side an output field that is expected to be output from the output port 3-2 when an optical signal is launched into the input port 3-1. According to the field distributions 1 and 2, a spatial refractive index distribution is calculated such that the phase difference between the propagation light and reverse propagation light is eliminated at individual points (x, z) in the medium.

The invention relates to a traffic sign detection and identification method based on pruning and knowledge distillation. The method comprises the following steps: preparing a data set and carrying outdata enhancement; establishing a network and training the network: establishing a YOLOV3-SPP network, loading parameters of a pre-training model trained in the data set ImageNet, and inputting the cut training set images subjected to data enhancement into the network in batches for forward propagation to obtain a model which is an original YOLOV3-SPP network; sparse training: using a scaling coefficient of a BN layer as a parameter for measuring channel importance, adding an L1 regularization item on the basis of an original target function, after adding the L1 regularization item, performingtraining again until loss convergence, and naming the process as sparse training; pruning according to the threshold value; and obtaining a final model by knowledge distillation.

The invention discloses a magnetic resonance reconstruction method based on deep learning and convex set projection, and relates to the technical field of magnetic resonance. The method comprises thesteps that S1, a network is constructed according to overlapping structures of multiple convolutional neural network modules and multiple convex set projection layers and shared data, wherein the shared data includes acquired K-space data and coil sensitivity information, and the convex set projection layers are obtained on the basis of the shared data; S2, after the network is constructed, all network parameters are trained through a reverse propagation process and verified; S3, structure and operation characteristics of the network are determined according to the verified network parameters,known test set data is input, the forward propagation of the network is conducted, unknown mapping data is obtained, and magnetic resonance reconstruction is completed. The method solves the problemthat by means of a current magnetic resonance reconstruction technology based on deep learning, only single-channel magnetic resonance data can be supported, and multi-channel magnetic resonance datacannot be processed.

The invention discloses a method for inverting a near-surface velocity model by utilizing preliminary waveforms. The method comprises acoustic wave equation-based wave field forward modeling and steepest descent-based waveform inversion technologies, and comprises the following steps of 1, extracting time-domain preliminary waveform records and an initial model; 2, calculating a simulated wave field and a wave field residual by utilizing acoustic wave equation staggered grid finite-difference forward modeling simulation; 3, reversely propagating the wave field residual to obtain a retransmission wave field; 4, calculating a gradient of a target function by utilizing the retransmission wave field and a forward propagation wave field, and calculating an updating step length; 5, updating a speed model; 6, inspecting whether the speed model is consistent with an iteration stopping condition, outputting the speed model if the speed model is consistent with the iteration stopping condition, otherwise returning to the step 2, and continuing iterative updating. According to the method, a wave equation theory-based full-waveform inversion technology is used as reference, and preliminary waves with higher energy and more stable waveforms are used for inversion, so that the multiplicity of solutions of full-waveform inversion is reduced, and the inversion stability and the calculation efficiency are improved; the accuracy of static correction and shallow depth imaging is improved.