167results about How to "Reduce network parameters" patented technology

The invention discloses a face identification method and apparatus based on a sequencing neural network model. The method comprises the following steps: performing preprocessing operation on input face images, and correcting angles and expressions of the face images; extracting features of already corrected face images / videos by use of a neural network comprising sequencing operation; and according to feature expression of the face images, calculating similarity between image pairs, and accordingly, obtaining identifies of specific objects in the input face images. According to the invention, a sequencing neural network structure is brought forward for solving the problems of too many parameters and too large calculation cost of a conventional neural network based face identification model in face identification problems, through sequencing expression between different features, network parameters are effectively reduced, and the calculation time is saved; and a training method based on comparison loss and triple loss is brought forward for solving the problem of a quite small number of training data.

The invention belongs to the field of image and video processing, and relates to a small target detection method based on multi-scale images and weighted fusion loss, and the method comprises the steps: extracting a plurality of groups of feature vectors from a plurality of different-scale images based on an improved Mask RCNN model, carrying out the fusion of the plurality of groups of feature vectors, and constructing a feature pyramid; generating a candidate detection box based on the feature pyramid and screening to obtain a suggested detection box; correspondingly returning the suggesteddetection boxes to the feature pyramid to generate feature maps of the suggested detection boxes, and performing aligned interception on the feature maps; inputting the aligned suggested detection boxes into a classifier layer to obtain category confidence coefficients and position offsets of the suggested detection boxes; in the test stage, screening a certain suggested detection box according tothe category confidence score of the suggested detection box, and performing non-maximum suppression; in the training stage, weighting the loss function calculated by detecting the small target feature layer and fusing with the loss function of detecting the large target layer and the middle target layer, thus the sensitivity of the model to the small target object is enhanced.

The patent discloses an intersection vehicle violation detection and identification method based on deep learning. The method comprises the following steps of 1, collecting intersection video detection images; 2, identifying traffic signs and vehicles; 3, tracking the vehicles; and 4, performing vehicle violation detection and identification. Based on automatic identification and tracking of the traffic signs and the vehicles in the intersection monitoring videos, the violation conditions of vehicles at the intersection can be automatically identified, and the system has few false detection and high efficiency.

The invention discloses a fundus image optic disc segmentation method and a system for assisting glaucoma screening, and relates to the technical field of image information processing. The fundus image optic disc segmentation method comprises the steps that a plurality of fundus images are collected and preprocessed, and a training image sample set and a verification image sample set are obtained;training of a constructed W-Net-Mcon full convolutional neural network by using the training image sample set to obtain an optimal W-Net-Mcon full convolutional neural networkis carried out; preprocessing the fundus image to be segmented, and inputting the preprocessed fundus image to be segmented into the optimal W-Net-Mcon full convolutional neural network to obtain a prediction target result image; Processing prediction target result graph by utilizing polar coordinate inverse transformation and ellipse fitting to obtain final segmentation result so as to obtain cup-to-disk ratio and finally obtain glaucoma preliminary screening result. According to the method, image semantic information can be effectively extracted in a multi-size mode, fusion of features of different levels, fusion of global features and detail features and encouragement of feature multiplexing are carried out, gradient back propagation is improved, and the image segmentation precision is improved.

The invention provides a Maxout multi-convolution neural network fusion face recognition method. The technical problem of unequal classification of positive and negative samples in a single network is solved. The method comprises the steps that a regression tree combination algorithm and affine transformation processing training data are used to acquire a database after alignment; face images in the aligned database are extracted to acquire a database simply with face area images; two networks are built, and a Maxout method is used to acquire a fused network; the database simply with face area images is used to train the fused convolution nerve network to acquire a trained network model; and a test image is pre-processed to test the trained network model. According to the invention, a Maxout module is used to fuse two or more subsystems, and then is successively connected with a full connection module and an output module to acquire a complete face recognition system; the classification sensitivity of positive and negative samples of the convolution neural network can be balanced; and the face recognition accuracy is improved.

The invention discloses a recursive residual attention network-based image super-resolution reconstruction method. The method is characterized by comprising the following steps of: 1) preprocessing data; 2) establishing a reconstruction model; 3) extracting features of a first residual attention module of a residual attention network branch; 4) extracting features of a first recursive module of arecursive network branch; 5) fusing the features; and 6) reconstructing an image. According to the method, noises caused by preprocessing can be solved, more high-frequency information can be obtainedto enrich the image details, the network parameters can be decreased, new parameters are not increased while the layers are increased; and the precision of super-resolution reconstruction can be improved.

The invention provides an infrared image weak and small target detection method based on improved YOLOv3. The method comprises the steps: firstly improving the standard convolution operation in a YOLOv3 residual module through deep separable convolution in a lightweight feature extraction network, then introducing a channel self-attention mechanism into each residual module of the lightweight feature extraction network, introducing a spatial self-attention mechanism into each residual module of the lightweight feature extraction network, finally, accelerating network training in the lightweight feature extraction network by using an H-switch activation function, further constructing an infrared image weak and small target detection model based on improved YOLO v3, and constructing a lightweight feature extraction network. In the network model design process, depth separable convolution is used for replacing the standard convolution operation of YOLO v3, different receptive fields are obtained through multi-scale feature map extraction, parameters are reduced, and therefore the method has the advantages that the network parameters are greatly reduced, and the calculation speed is high.

Image pairing information and image classification information are optimized and a Hash code quantization process is used to realize a simple and easy end-to-end deep multi-scale supervision Hash method, and meanwhile design a brand new pyramid connected convolutional neural network structure, and the convolutional neural network structure takes paired images as training input and enables the output of each image to be approximate to a discrete Hash code. In addition, the feature map of each convolution layer is trained, feature fusion is carried out in the training process, and the performance of deep features is effectively improved. A neural network is constrained through a new binary constraint loss function based on end-to-end learning, and a Hash code with high feature representationcapability is obtained. High-quality multi-scale Hash codes are dynamically and directly learned through an end-to-end network, and the representation capability of the Hash codes in large-scale image retrieval is improved. Compared with an existing Hash method, the method has higher retrieval accuracy. Meanwhile, the network model is simple and flexible, can generate characteristics with strongrepresentation ability, and can be widely applied to other computer vision fields.

The invention discloses a quick magnetic resonance imaging method based on a recursive residual U-type network. The method comprises the three steps of data processing, model training and image reconstruction. By adopting the method, a recursive residual module is added to the U-type network, the problems of gradient blast and gradient vanishing caused by back propagation of the network are solved, new network parameters are not added while the layer number of the network is increased, the quality of a reconstructed image is obviously improved, and more image details can be restored.

The invention relates to an attention mechanism-based image aesthetics quality evaluation method, and belongs to the technical field of computer vision. The method comprises the steps of firstly processing training data, then designing a network structure model, adopting a lightweight deep network as a backbone network, and integrating the backbone network into an attention mechanism module; designing a loss function for training the network based on a data equalization thought; finally, training a network structure model by using the processed training data to obtain a network model capable of automatically evaluating the aesthetic quality of an image; and performing aesthetic scoring on the input picture based on the model, and applying the model to shooting to assist a user to shoot a more beautiful picture in real time. Compared with the prior art, the network structure model adopted by the invention can more effectively extract the characteristics of images, the adopted loss function greatly enhances the data learning ability of the model. Compared with other methods, the accuracy is improved, and the parameter quantity of the model is reduced.

The invention discloses a substation patrol robot road scene recognition method based on depth learning. The method includes collecting substation road scene image and constructing road scene image database, constructing convolutional coding network and training, constructing deconvolution decoding network and training based on convolutional coding network, and using test set to test model and road scene recognition application of substation patrol robot. The whole convolutional neural network constructed by the invention simultaneously gives attention to identification accuracy and efficiency, reduces network parameters and storage space by simplifying the deep high-precision convolutional neural network into a small shallow network to improve identification efficiency, and adopts a plurality of decoding network fusion modes to obtain more target information to ensure identification accuracy. The depth learning scene recognition method can obtain the dense information of the current environment of the substation inspection robot, provide more effective guidance information for robot navigation obstacle avoidance, and make the robot more adaptable to the environment.

The invention discloses a pedestrian re-identification method based on natural language description and relates to processing of a recording carrier for identifying graphics. Specifically, the invention discloses an image and a natural language are designed to describe a double-branch network structure. In the image branch network structure a MobileNet convolutional network is adopted to carry out image feature extraction to extract text features of a natural language description branch network structure through a BiLSTM network. a stacking loss function is created for similarity measurementparts between image features and text features, and network training is carried out. A corresponding pedestrian image contained in the to-be-detected image set is searched by using a trained network to achieve pedestrian re-identification based on natural language description of a stacking loss function . The defects that in the prior art, the text feature characterization of a feature extractionpart is not high, a loss function part is difficult to train a network for a long time, and a large amount of memory is consumed in the training process are overcome.

The invention provides a driving fatigue detection and early warning system based on vision. The driving fatigue detection and early warning system comprises a face detection and identity verificationmodule, a normal driving baseline database module, a fatigue driving behavior detection module, a fatigue driving behavior early-warning module and a fatigue driving behavior management and control module; the face detection and identity verification module is used for collecting images of a driving cab and extracting face information and identity information; the normal driving baseline databasemodule is used for establishing and updating fatigue baselines of a driver; the fatigue driving behavior detection module is used for generating fatigue driving ROI based on the extracted face information, and conducting fatigue driving behavior identification; the fatigue driving behavior early-warning module is used for judging the driver fatigue state and the driving fatigue level according tothe monitoring results of the fatigue driving behavior detection module and issuing a fatigue early-warning signal for the driver; and the fatigue driving behavior management and control module is used for triggering driving data records and uploading data according to the fatigue early-warning signal.

The invention discloses a human posture recognition method based on depth learning, which mainly solves the problems of large calculation amount and low accuracy of the current posture recognition technology. Firstly, the kinect V2.0 depth sensor is used to collect the motion and posture characteristics of several human body samples; RGB data and bone data of human motion posture are stored; skeletal images obtained from bone data after image preprocessing are used as training set and test set; the training set is inputted into a Posture based on Convolution Neural Network (CNN), which is specially used in the field of human posture recognition. In CNN, the classification result is obtained after training and adjusting the network structure and network parameters. The motion and posture characteristics of different human body samples are input into the classification network as a test set, and the most probable motion is the recognition result. The invention uses the convolution neuralnetwork to improve the identification accuracy, reduce the identification time, has low operation cost, and is simple and convenient to be applied in places such as intelligent home, safety monitoring, motion analysis and the like.

The invention relates to a fast light super-resolution reconstruction dense residual error network. Based on a dual-channel deep residual error network (FLSR) of a convolutional network, a deep channel is mainly used for learning high-frequency texture information of an image, and a shallow channel is used for learning low-frequency information of of the image. In order to accelerate the convergence speed of the network, a residual error connection mode is added in the structure, in addition, the image detail information of the front convolution layer can be directly transmitted to the rear convolution layer through the residual error connection mode, and reconstruction of images with better quality is facilitated; and a dense connection mode which contributes to weakening gradient disappearance and improving the model performance is also adopted in the structure. While image quality evaluation indexes such as peak signal-to-noise ratio (PSNR), structural similarity (SSIM) and information fidelity (IFC) standard measurement are improved, parameters and computational complexity of the model are reduced, the reconstruction speed of the image is increased, and the method can be applied to actual life.

The invention discloses a multi-objective fused educational resource personalized recommendation system, which is characterized by comprising a recommendation system, and the recommendation system comprises a service layer, a recommendation layer, a strategy layer and a data layer. The data layer comprises data displayed and fed back by a user, a user interest model, resource content and overall resource content. The strategy layer comprises representation features of learner users and representation features of education resources. The recommendation layer comprises an SOM-CNN model, an ITEM-SOM model and a recommendation list generated according to the SOM-CNN model and the ITEM-SOM model. The service layer comprises education resource recommendation services for different users. The idea of integrating multi-task target learning is applied to the educational resource recommendation system, personalized recommendation of educational resources is realized, the recommendation accuracyand diversity are used as two to-be-learned target tasks of the network, and two different network module structures are adopted to concentrate on the accuracy and diversity of recommendation resultsrespectively. And a unified loss function is designed, and synchronous end-to-end learning and training are performed on multiple targets of the network recommendation result.

The invention relates to an expression recognition method, in particular to an expression recognition method based on a multi-branch cross-connection convolutional neural network. The invention aims to solve the problems of low efficiency, serious resource waste and incomplete feature extraction of an existing traditional expression feature extraction method. The method comprises the following steps of: 1, preprocessing a facial expression image data set; 2, a multi-branch cross-connection convolutional neural network is constructed and used for extracting facial expression image features, andthe process is as follows: the multi-branch cross-connection convolutional neural network is composed of a first convolutional layer, a module 1, a module 2, a module 3, a forty-th convolutional layer, a batch standardization BN and a Relu activation function; and 3, classifying the image features extracted by the network by adopting a Softmax classification algorithm, namely connecting a globalmean value pooling after the constructed multi-branch cross-connection convolutional neural network, and carrying out multi-classification by using a Softmax function after a global mean value poolinglayer. The method is applied to the field of expression recognition.

The invention discloses a fixed subarray space-based millimeter wave beam forming method based on a depth complex network, which simplifies a radio frequency pre-coding matrix and a radio frequency merging matrix through a fixed subarray structure, greatly reduces the requirements of a phase shifter, reduces the hardware dependence and reduces the failure rate compared with the prior art. By constructing the neural network which inputs the real part and the imaginary part of the complex number respectively and is fully connected with the lower-layer neurons, compared with a traditional methodin which the real part and the imaginary part need to be separated into two networks for processing, computing resources are saved; through the design of weight and offset parameter shared imaginary part neuron and real part neuron node expressions, half of network parameters are reduced, the system complexity is greatly reduced, hardware resources are saved, the operation rate is increased, and the real-time performance is improved.

The invention provides a method for performing human body behavior recognition by using an attitude mask, and belongs to the technical field of behavior recognition. The method comprises: extracting athermodynamic diagram from the frame image of the RGB video by using a preposed two-dimensional attitude estimation network; and performing multi-point Gaussian diffusion on the thermodynamic diagramto obtain an attitude mask, taking an image sample obtained by fusing the attitude mask and the original image as input of a space-time neural network, and training the space-time neural network by using the training set marked with the human body behavior category label to obtain a model for human body behavior recognition. According to the method, the spatial features of the human body skeletonkey point position area obtained through attitude estimation are extracted through the attitude mask, the robustness to the change of the image background is high, the number of recognition network parameters is small, the training cost is low, and the recognition accuracy of human body behaviors is high.

The invention provides an urban sound event classifying method based on an N-DenseNet and high-dimensional mfcc feature. By the urban sound event classifying method, during audio data processing, richer and more efficient feature information can be provided, ta model has stronger generalization ability, and classification has higher accuracy. The urban sound event classifying method comprises thefollowing steps: S1, acquiring audio data to be processed, preprocessing an original audio signal, and outputting an audio frame sequence; S2, performing time-domain and frequency-domain analysis on the audio frame sequence, extracting a high-dimensional Mel-frequency cepstrum coefficient, and outputting a feature vector sequence; S3, constructing an acoustic model, and training the acoustic modelto obtain a well-trained acoustic model; S4, processing the feature vector sequence output in the step S2, and then inputting into the well-trained acoustic model for classification recognition to obtain a recognition result, namely a classification result of a sound event, wherein the acoustic model is a network model constructed by combining the characteristics of an N-order Markov model on thebasis of a DenseNet model, namely the acoustic model is an N-order DenseNet model.

The invention discloses a large-scale image sub-block retrieval method based on a deep Hash network and sub-block reordering. The large-scale image sub-block retrieval method comprises the following steps: preparing a deep network weight pre-trained by a large-scale image database and a texture picture library to be trained; extracting picture pairs from the texture picture library, when the two pictures have the same label, determining the two pictures as positive samples, and taking the two pictures with different labels as negative samples; inputting the picture pairs into two same shared weight networks for training in pairs, and setting a loss function to binarize a network output result; quickly obtaining similar pictures by using the network as a Hash code extractor; after subblockreordering is conducted on the similar pictures, quickly mapping subblocks in an original picture and obtaining a large number of similar sub-blocks. According to the method, the calculated amount isgreatly reduced, rapid retrieval of large-scale image sub-blocks is achieved, and the method can be applied to various image enhancement methods.

The invention discloses a convolutional neural network based SAR moving target indication method, and is applied to the field of SAR moving target indication. For defects in the prior art of high requirements for the number of available to-be-detected target auxiliary data range gates, the method calculates target speed by constructing a convolutional neural network and training the neural networkaccording to the total Doppler frequency obtained through the detection of the neural network and angles of target relative synthetic apertures, so that SAR moving target indication can be realized;and the method can realize detection under the circumstance of a small number of the available to-be-detected target auxiliary data range gates.

The invention discloses an MCS (Mesoscale Convection System) identification and tracking method based on image anchor-frame-free detection, which comprises the following steps of: step 1, preprocessing infrared brightness temperature data of an original stationary satellite, carrying out mesoscale convection system marking on an infrared cloud picture obtained after processing, and then randomly dividing a training set, a verification set and a test set; step 2, constructing an instance segmentation network based on no anchor frame, the network being used for extracting image features, detecting a mesoscale convection system and segmenting specific instances; step 3, performing training set image enhancement, using a transfer learning supervised training instance to segment the convolutional neural network, and automatically learning network parameters; step 4, performing mesoscale convection system detection and segmentation on the geostationary satellite infrared nephograms at adjacent moments by using the trained model; and step 5, realizing the tracking of the mesoscale convection system according to a related target matching principle.

The invention provides a three-dimensional face and eyeball movement modeling and capturing method and system, and the method comprises the steps: firstly extracting the position information of a feature point of a face in a shot scene image, and obtaining the information through employing a convolutional neural network method; after sparse facial feature points are obtained, a multilinear model of a human face serving as a prior condition of the model, the positions of the feature points detected on the image serving as observation results, and designing a maximum posteriori framework to optimize and solve the geometrical shape and posture of the human face in the image; aiming at the movement of a three-dimensional eyeball, using sparse two-dimensional feature points for obtaining an eyearea image block through cutout, achieving alignment of the image block through position information of the feature points, further marking the areas of an iris and a pupil, and reconstructing the eyeball movement in real time through an analysis method based on synthesis; in combination with the expression parameters obtained in the previous step, combining facial expression motion with eyeballmotion to obtain a complete facial expression animation.

The invention discloses a method and device for obtaining the feature points of a human face. The method comprises the steps: obtaining a human face after normalization processing; inputting the humanface into a convolution neural network for processing, wherein the convolution neural network comprises at least one convolution layer, at least one pooling layer, at least one local response normalization layer and at least one total connection layer, and the convolution layer is used for the convolution processing according a convolution core, the pooling layer is used for simplifying the dataof the convolution layer, and the configuration of the convolution neural network is correlated with the expected number of feature points of the human face, and is obtained through training via a predetermined training set; obtaining a plurality of feature points, processed by the convolution neural network, of the human face in the human face image. The method solves technical problems that a method for obtaining the feature points of the human face based on the convolution neural network in the prior art is large in calculation burden, and is longer in training time.

The invention relates to the technical field of water turbine runner blade defect detection, and particularly discloses a water turbine runner blade defect detection method based on a YoloV4-Lite network, and the method comprises the steps: S1, constructing a defect detection network based on the YoloV4-Lite network; S2, performing picture collection on different defect positions of the turbine runner blade to obtain thousand or more defect pictures; S3, preprocessing the defect picture acquired in the step S2 (processing by using LabelImg software according to a Pascal VOC 2012 format) to obtain a data set; and S4, training, testing and verifying the defect detection network by adopting the data set. According to the invention, the backbone extraction network CSPDarkNet53 network of YoloV4-Lite is replaced by the MobileNet network, and the MobileNet network is a real-time lightweight network, so that the network detection speed can be improved, and the network parameters can be greatly reduced. Experimental results show that the accuracy rate of the defect detection network can reach 97.48%, the network parameter quantity of the MobileNetV3 only needs 37.35 MB and is reduced by 206.94 MB compared with that of CSPDarkNet53, the FPS reaches 44.68, and the method has the advantages of being high in accuracy rate, low in memory storage and real-time.

The invention discloses a magnetic resonance imaging method and device. According to the method, after a first deep neural network is used for reconstructing a first image, a first part of k-spatial data and a second part of k-spatial data extracted from complete k-spatial data are subjected to subtraction operation to obtain residual k-spatial data; then the residual k-spatial data is subjected to image reconstruction to obtain a residual image; and finally, the first image and the residual image are added to obtain a magnetic resonance image. In this way, the first part of the really collected k-spatial data is utilized in the processing carried out after the first image is reconstructed, the processing process of data fidelity terms is carried out outside the DNN reconstruction, and theDNN does not need to consider the processing of the data fidelity terms, so that the complexity and network parameters are lowered and reduced, and the image reconstruction speed is high. In addition, the processing process of the data fidelity terms is completed through sparse constraint reestablishing, without multi-time iteration, and the reconstruction speed can be improved on the premise that the image reconstruction stability is improved.

Because the generalization of a flame detection model based on image features is not strong, and the requirement of a deep neural network model for the number of training samples is high, the invention provides a flame target detection method based on digital images and convolution features, and the method comprises the steps: firstly making a data set comprising video dynamic features; replacingthe standard convolution of the VGG16 in the classic Faster R-CNN with the depth separable convolution, and reducing the number of convolution layers; cutting 256 image blocks from the original imageaccording to a candidate box generated by the RPN, and extracting LBP features of each image block; reducing the size of an output feature map of the ROI pooling layer and the number of neurons of a full connection layer through convolution, and further reducing network parameters; and finally, combining the extracted LBP features, the dynamic features in the data set and the pooled tiled featurevectors, and sending the combined feature vectors to a full connection layer for classification and regression. The flame target detection model constructed by the patent has relatively high detectionprecision, is convenient to improve for overcoming the defects of a test result, and is high in flexibility.

The invention discloses a method of recommending a personalized treatment scheme for a stroke patient. The method comprises the following steps: S1, preprocessing text information about physical examination and evaluation results in electronic medical records of patients; S2, expressing words, sentences and documents in the physical examination and evaluation results in the electronic medical records of the patients in a vector manner; S3, training a neural network model based on document vectors to obtain a personalized treatment scheme recommendation model; and S4, carrying out unified dataexpression, word segmentation and text filtering processing on the physical examination and evaluation results in an electronic medical record of a new patient, then carrying out document vector representation, and inputting represented document vectors into the personalized treatment scheme recommendation model to obtain a recommended personalized treatment scheme. According to the method, evaluation and physical examination information in the electronic medical record of the patient is taken as a document, the process of personalized treatment scheme recommendation is converted into a multi-label classification problem, the personalized treatment scheme can be recommended according to the physical examination results and the evaluation results of the patient, an auxiliary decision is provided for a doctor, and the burden of the doctor is reduced.

The invention provides a microalgae identification method based on improved YOLOv3, which comprises: collecting microalgae microscopic images, and making a data set of the microalgae images; performing data enhancement on the data set; dividing the enhanced data set into a training set, a verification set and a test set, labeling microalgae in the data set, and generating a labeled image; constructing an improved YOLOv3 target detection model; setting training parameters, and training the constructed YOLOv3 target detection model based on the data set; and classifying and positioning the test set images based on the trained YOLOv3 target detection model. According to the method, an improved YOLOv3 target detection model is adopted, a lightweight Mobilenet network is used for replacing an original feature extraction network darknet53 of YOLOv3, the operation speed can be remarkably increased, network parameters are greatly reduced, meanwhile, a spatial pyramid pool structure SPP is introduced, region features can be combined and connected in the same convolution layer with different scales, and the method is suitable for large-scale detection, so the position error is small when a small object is detected, and the CIoU is used for optimizing the loss function to further improve the detection precision.