329 results about "Residual neural network" patented technology

The invention discloses a deep-convolution-neural-network-based CT pulmonary nodule detection method. The method comprises: (1), CT image pretreatment is carried out, so that the pixel spacing becomes unified and image comparison is unified; (2), a two-dimensional convolution neural network U-net is trained, a pulmonary nodule segmentation image is predicted, and a candidate nodule is recommended based on the pulmonary nodule segmentation image; and (3), a three-dimensional deep residual neural network Resnet3D is trained, a true-false positive probability of the pulmonary nodule is predicted, and a false positive nodule is screened out. According to the detection method disclosed by the invention, the deep learning advantages are utilized fully, so that the pulmonary nodule can be detected in a CT image automatically, efficiently and accurately; and the adaptability to medical big data is high.

The invention discloses a video compression method based on the deep neural network. The method includes the following steps that a video image data set is collected and organized, and a neural network training set, a test set and a cross validation set are constructed; the multi-layer deep neural network is set up; for inter-frame prediction, a motion estimation algorithm is used for searching for the optimal matching block, and residuals and the mean square error of inter-frame prediction are calculated; the residuals obtained after prediction are used as new training data to train a residual code network, and a residual network model comprises an intra-frame residual and an inter-frame residual; compression data serving as fixed-length code streams together with output data, obtained after quantization and lossless entropy coding, of the residual neural network is predicted; a decoding terminal restores the compression data through a neutral network symmetric with a coding terminal, and a compressed image is obtained through reestablishment and recovery. Compared with a traditional H.264 video coding method in equality comparison of plenty of test video sequences, the video compression method can save about 26% code rate on the premise of equal quality.

The present invention relates to an image fusion method, in particular to a multimodal image fusion method, especially a multi-modal image fusion method based on a generative adversarial network and asuper-resolution network. The method is carried out according to the following steps of designing and constructing the generative adversarial network, wherein the network structure adopts a conceiveddepth residual neural network, and obtaining a generating model through the dynamic balance training of a generator and a discriminator; constructing the super-resolution network based on the convolution layer; inputting the multi-band / multi-mode source image into the generating model to obtain the preliminary fusion image; and then inputting the image into the trained super-resolution network toget the final fusion image with high quality. The method realizes the end-to-end neural network fusion of multi-band / multi-mode images, avoids the difficulties of image multi-scale and multi-direction decomposition and fusion rule design based on prior knowledge, and realizes the adaptive network fusion.

The invention relates to a three-dimensional scene reconstruction method based on depth learning, belonging to the technical field of depth learning and machine vision. The depth structure of the scene is estimated by convolution neural network, and the dense structure is refined by multi-view method. A full convolution residual neural network is trained to predict the depth map. Based on the color images taken from different angles, the depth map is optimized and the camera attitude is estimated by using epipolar geometry and dense optimization methods. Finally, the optimized depth map is projected to three-dimensional space and visualized by point cloud. It can effectively solve the problem of outdoor 3D reconstruction, and provide high-quality point cloud output results; the method canbe used in any lighting conditions; It can overcome the shortcoming that the monocular method can not estimate the actual size of the object.

The invention discloses a text recognition method based on attention mechanism. The network SAN based on spatial attention is an end-to-end text recognition model. The text recognition model comprisesa feature extractor with a local neural network, a residual neural network and coordinate information, and a spatial decoder based on attention mechanism. The text recognition model is based on the encoding and decoding structure, so the text recognition model can also be understood as encoder and decoder. The encoder is used to encode the input image to obtain the encoded feature sequence whichcan be recognized by the decoder. A decoder is used to decode the encoded features of the encoder, thereby realizing recognition of text in an image. For the arc text CUTE80 dataset, the result of this method is superior to all the existing methods, and the accuracy is 77.43%. In other scene text datasets, this method also has a good effect.

A CT image pulmonary nodule detection method based on a 3D residual convolutional neural network includes a training process and a testing process. The training process includes the following steps: S1, preprocessing an original image, resetting the voxel spacing to (1, 1, 1), and converting the voxel spacing into voxel coordinates; S2, capturing 3D positive and negative samples from a CT image; S3, setting a maximum and a minimum, and standardizing the sample data; S4, constructing a 3D convolutional neural network; S6, setting training hyper-parameters, and importing the training hyper-parameters to a data training model in the form of mini-batch; and S6, saving the model after the model is fully trained. The testing process includes a step S7: preprocessing test CTs, sampling the test CTs one by one in the form of sliders, importing the test CTs to the model for calculation, selecting samples with high confidence, and deleting repeated samples through a non-maximum suppression algorithm. The method is of high accuracy, and can be used to analyze whether there is a nodule in an image and the specific position of the nodule in the image.

A learned image compression system increases compression efficiency by using a novel conditional context model with embedded autoregressive neighbors and hyperpriors, which can accurately estimate the entropy rate for rate distortion optimization. Generalized Divisive Normalization (GDN) in Residual Neural Network is used in the encoder and decoder networks for fast convergence rate and efficient feature representation.

The present invention relates to the technical field of image recognition. A multi-view pointer instrument identification method includes: acquiring an image and uploading the image to a computer; using a SSD algorithm to position an instrument area; using a ResNet34 deep residual neural network to perform classification training on the instrument area, and performing preliminary correction on thesample image according to a classification result; using the SSD algorithm to secondarily position the instrument area of the corrected image; using the network to perform regression training on thesecondarily positioned instrument area to identify the position of a pointer on an instrument dial; using an HED edge detection algorithm to perform dial edge detection on the positioned instrument area; performing random sampling according to an RANSAC algorithm, and calculating the edge model of the instrument dial; correcting the instrument point by a zoom ratio, and calculating the angle between the instrument point and a start point; and looking up a database to obtain the scales of the instrument dial. The method can recognize the pointer instrument captured at different angles.

The invention relates to a photovoltaic cell appearance defect classification method based on a multi-channel residual neural network. The method classifies the photovoltaic cell appearance defects based on a depth learning algorithm of a multi-channel input residual neural network. Firstly, the acquired photovoltaic cell sheet appearance image is preprocessed. 20% of that target image are randomly selected as a t sample set, the remaining target images are manually sorted, label are added, the size of the target images is quantized and multi-channel information in the target images is extracted, so that the training sample sets with fixed scales are obtained respectively, and the sample sets are verified. The training set is inputted into the residual neural network, and the multidimensional output eigenvalue matrix of the image is obtained. According to the extracted multi-dimensional eigenvalue matrix, the verification set image features are loaded into softmax classifier for classification, and the classification results are compared with the labels, and the test data and multi-dimensional eigenvalue matrix are loaded into the classifier to obtain the final classification. Thisapplication has high accuracy and high speed.

The invention relates to a non-contact atrial fibrillation intelligent detection system based on a deep convolution residual network. Heart impact signals generated during cardiac ejection of the human body can be collected, and whether or not the heart has atrial fibrillation is analyzed by using the deep convolution residual network. The system includes two parts including a heart impact signalcollection part and a deep convolution residual neural network analysis part. The heart impact signal collection part includes a piezoelectric film sensor, a piezoelectric signal processing module, anAD conversion circuit and a processor. Collected data is uploaded to a server running a deep convolution residual network algorithm, and a judgment result is turned back. A deep convolution residualneural network model includes a one-dimensional convolution layer, a downsampling layer, a batch normalization layer, an activation layer and a fully connected layer. A residual structure contained inthe model is formed by adding original signals and deep features extracted after several convolutional sample normalization activation operations. The system can be used for uninterrupted monitoringof the human body at night and give a judgement prompt in time to facilitate timely treatment of disease conditions.

The invention discloses a steel rail scale damage detection method based on deep learning. The method comprises the following steps: firstly, dividing a data set formed by all images into a training set, a test set and a verification set according to a set proportion; setting a network structure and performing forward propagation, and performing deep learning training by using a residual neural network, the residual neural network comprising a convolutional layer, a pooling layer and a full connection layer; after a calculation result of forward propagation is output, calling a reverse propagation algorithm; and finally, reserving a model trained by the final residual neural network, and drawing a change curve of each parameter in the whole training process for reference. According to themethod, a convolutional neural network technology in machine vision and deep learning is combined, features of a steel rail scale damage sample are extracted and learned and classified, and a model output by the neural network is used for judgment in the actual industry; compared with a method for judging the fish scale damage on the surface of the steel rail by using a manual method in the industrial field, the method has very high detection efficiency and accuracy.

The invention discloses an electrocardiosignal detection device and analysis method based on a joint neural network. The method comprises the following steps: firstly, building a joint neural networkalgorithm on a machine learning server, and training a model; aiming at preprocessed ECG data, enabling the model to extract data spatial features and acquire a spatial classification probability through a residual neural network module; extracting time sequence features of the data on a dimensionality-reduced spatial feature map through a bidirectional long-short-term memory neural network and anattention module, and acquiring a time sequence classification probability; finally, fusing the two classification probabilities to obtain a detection result; acquiring a small amount of ECG data ofa patient from a wearable device, performing manual marking, inputting the ECG data into the machine learning server, performing fine-tuning on the model, and deploying the final model to an intelligent mobile device; and finally, realizing real-time anomaly detection through wireless transmission of the wearable device and the intelligent mobile device. The invention develops the wearable devicefor electrocardiosignal acquisition and the real-time detection, and provides an effective technical means for auxiliary diagnosis of heart diseases.

The invention discloses a gesture recognition method based on an improved residual neural network. The method includes the following steps: S1, acquisition of a training sample set; S2, preprocessingon the training sample set, wherein positions of gestures in images are found through algorithms, and cropped images are used as original training data; S3, enhancement of training samples, wherein translation transformation, rotation transformation, mirror-image transformation, scaling transformation and the like are carried out on the collected training samples to enlarge the training sample set; S4, acquisition of a gesture model, wherein a processed training sample set is input into the pretrained residual network to carry out training on network parameters to obtain the gesture recognition model; S5, a step of carrying out processing, which is the same as the step S2, on to-be-recognized gesture images to obtain to-be-recognized gesture data; and S6, a step of inputting the to-be-recognized gesture data into the network, of which training is completed, to obtain a gesture sequence. The method is based on the deep residual network, trains the residual network on the self-collecteddata set, and realizes high-recognition-rate gesture recognition of a third view angle.

The invention discloses a channel estimation method for a passive intelligent reflection surface based on deep learning. The channel estimation method is realized by an offline channel estimation stage and an online channel prediction stage. In the off-line channel estimation stage, in an uplink, a user side sends a pilot signal, a base station side controls an IRS to sequentially open passive elements to reflect an incident pilot signal, and the base station side receives the pilot signal and estimates corresponding cascade channel information through adoption of a minimum mean square error method. An equal probability uniform sampling method is adopted to select a small amount of sampling cascade channel information from the estimated cascade channel information, and the small amount of sampling cascade channel information and complete cascade channel information are adopted to construct a new data set; and in the online channel prediction stage, the base station side estimates a small amount of sampling cascade channel information online and inputs the sampling cascade channel information to the trained ResNet network to recover complete cascade channel information. According to the invention, the number of passive elements can be flexibly selected and the residual units of the residual neural network can be set so as to meet the service quality characteristics of different systems and users.

The invention discloses a protein-ligand binding site prediction algorithm based on deep learning. For protein to be predicted, the algorithm comprises the steps of: firstly, extracting sequence features and a distance matrix; distributing the sequence features to each residue through adoption of a sliding window method; and inputting the features corresponding to the residues into a residual neural network and a hybrid neural network one by one, and inputting output results of the residual neural network and the hybrid neural network into a Logistic regression classifier to obtain a final result, namely the binding probability corresponding to each residue in the protein. According to the method, a classic bidirectional long-short-term memory network and a residual neural network are fused, the fused network can process heterogeneous protein sequences and structural data at the same time, and complementarity of sequence features and structural features is mined. Compared with an existing method, the protein-ligand binding site prediction algorithm has higher prediction precision, and has good generalization performance for data sets of different ligands.

The invention discloses a Bayesian wide-residual neural network-based injurious insect image recognition method. The method comprises the following steps of: 1, preprocessing an injurious insect imagerecognition data set and carrying out injurious insect edge detection on a greyscale map by using a Rich-Edge detection algorithm; 2, constructing a Bayesian wide-residual neural network (NWResNet);3, inputting an injurious insect edge image obtained in the step 1 into the NWResNet constructed in the step 2, and obtaining a total error function of the NWResNet by utilizing an injurious insect edge image training set obtained in the step 1; 4, training the network through a Block-cg algorithm by utilizing the error function obtained in the step 3; 5, updating a hyper-parameter according to the network optimized in the step 4; and 6, repeating the steps 4 and 5 to obtain a final network, and inputting an injurious insect image verification set into the network to ensure that the classification correctness is higher.

The invention discloses an electric vehicle charging navigation method based on prediction of the dynamic occupation rates of charging piles in charging stations. Aiming at the problems that an electric vehicle is difficult to charge and the charging piles cannot be fully utilized, based on a deep residual neural network, the dynamic occupation rates of the charging piles in the charging stationsare predicted, the optimal charging pile in the charging stations is recommended to a user, and the path with the shortest consumed time is provided. Firstly, through a starting point set by the userand current electric quantity information of the electric vehicle, a background calculates to judge whether charging is need or not, if yes, all the charging stations capable of being reached of an area within a driving mileage are obtained, the deep residual neural network is adopted to predict driving-in and driving-away conditions of the vehicles in the charging stations, the occupation rates of the charging piles in the charging stations are calculated, based on a crowd sensing technology, the occupation rates are corrected in real time, and through the further utilization of the distancesfrom the starting point to the charging stations and the distance from the charging stations to an end point, the charging pile scheme with the optimal path is intelligently recommended to the user.

The invention discloses a remote sensing image fusion method and system based on a deep residual neural network, and the method comprises the steps of carrying out the downsampling of corresponding multiples on a panchromatic image and a multispectral image, and obtaining a training sample; constructing a deep residual neural network, wherein the network is of a deep convolutional structure, and inputting a panchromatic image and a multispectral image into the deep residual neural network to sequentially pass through outputs obtained by a plurality of convolution modules and adding with the input multispectral image to form a residual structure; using a training sample and a random gradient descent algorithm for training a deep residual neural network, downsampling a multispectral image and a panchromatic image to be fused respectively by corresponding multiples, then inputting the multispectral image and the panchromatic image into the trained deep residual neural network, and obtaining a fused multispectral image with high spatial resolution. According to the invention, the respective key information of the two images can be integrated, and the spatial resolution of the multispectral image is improved.

The invention relates to a forged face video detection method and device. The method comprises the following steps of: extracting a video frame sequence from the video data; inputting the video framesequence into a pre-trained residual neural network, extracting residual features corresponding to the video frame sequence; and inputting the video frame sequence into a pre-trained edge texture neural network, extracting edge texture features in the video frame sequence, performing feature fusion on the residual features and the edge texture features to obtain fusion features, inputting the fusion features into a pre-trained binary convolutional neural network, and outputting a detection result whether a face is a forged face or not. By adopting the method, the accuracy of counterfeit face detection can be improved.

The invention discloses an image restoration method based on an adaptive residual neural network. The method comprises the steps that an adaptive residual neural network model is built, wherein the adaptive residual neural network comprises a number of adaptive residual units connected in series; training sets for image denoising, image super-resolution and image deblocking effects are respectively selected, and corresponding training parameters are respectively set; according to the adaptive residual neural network model and the training parameters for image denoising, image super-resolution and image deblocking effects, the corresponding target neural network model is trained respectively with the goal of minimizing a loss function; and according to the trained target neural network model for the problem of image denoising, image super-resolution and image deblocking effects, an image to be processed is input to the corresponding target neural network model, and a corresponding high-quality image is output. According to the invention, the PSNR, SSIM and visual effect of the image can be remarkably improved, and the method has the advantages of good recovery effect, high speed and strong robustness.

The invention provides an expressway vehicle detection and multi-attribute feature extraction method based on a local image, and relates to the technical field of intelligent transportation. A video acquisition terminal reads expressway monitoring video in real time and transmits the expressway monitoring video to an edge end, and the edge end analyzes the real-time video data by adopting a background difference method to select a key frame; a cloud end uses a VOC2007 data set and vehicle pictures collected by an expressway to train a YOLO-v3-tiny detection model, the edge end loads the trained YOLO-v3-tiny detection model to predict the position of a vehicle bounding box in the selected key frame, and then a local image of a vehicle is obtained and transmitted to the cloud end; a ResNet-50 residual neural network model is trained by the cloud end by utilizing the training set data with the multi-label type, the edge end loads the trained ResNet-50 residual neural network model, and the acquired local image of the vehicle is input into the neural network model to realize the extraction of multi-attribute features of the vehicle; and the extracted multi-attribute features of the vehicle are made into a label, and the label is uploaded to the cloud end.

The invention relates to an electrocardiogram classification method of deep residual neural network based on an attention mechanism. According to the invention, automatic classification of arrhythmia is realized by multi-lead ECG data and a deep residual neural network model based on the attention mechanism, firstly, the multi-lead ECG contains more ECG information than the single-lead ECG, and secondly, the depth residual network can learn characteristics of the model in a higher mode, which helps the model to converge, and finally, an attention mechanism module automatically enhances a feature map, improves the generalization ability of the model, and improves the learning efficiency of the network and the accuracy of ECG recognition.

Embodiments of the invention provide a recursive residual network-based super-resolution image reconstruction method. The method comprises the following step of: inputting a low-resolution image intoa trained recursive residual neural network so as to obtain a super-resolution reconstructed image, wherein the recursive residual neural network comprises a plurality of residual units, and for any residual unit, input information of the residual unit comprises output information of the last residual unit and a high-frequency feature image of a low-resolution input image. According to the method,the neural network is trained through local residual learning but not global residual learning of VDSR, so that benefit is brought to information transmission and gradient flow; non-interpolated low-resolution images are taken as inputs; and finally, super-resolution output images are directly up-sampled by using a convolution layer at the tail end of the network, and recursive structures are imported in the residual units, so that parameters are greatly decreased and the calculation complexity of the recursive residual neural network is reduced.

The invention discloses an image description generation method and device based on a deep residual network and attention. According to the method, the problem that the precision of the deep neural network is reduced is solved, the deep residual network is used for learning the image features of the image from the bottom layer to the top layer, rich input image representations are generated, and then natural and smooth description sentences are generated in combination with the attention circular long-short-term memory network. The method comprises the following steps: acquiring a large amountof image sample data, preprocessing the image sample data; extracting image features of the preprocessed image sample data; processing the extracted image features by using a residual neural network model to generate an image representation; mapping the image representation to an input of an attention-based cyclic long-short-term memory network language model, predicting word vectors of the imageby using the attention-based cyclic long-short-term memory network language model, and generating a descriptive sentence of the image.

The invention discloses a semantic segmentation method based on a residual pyramid pooling neural network. The method comprises the following steps: firstly, constructing a convolutional neural network, wherein a hidden layer of the convolutional neural network comprises 10 residual neural network blocks, 4 Residual ASPP blocks and 5 Basic blocks; inputting the original indoor scene image into a convolutional neural network for training to obtain a corresponding semantic segmentation prediction graph; calculating a loss function value between a set formed by semantic segmentation prediction images corresponding to the original indoor scene images and a set formed by 40 single-heat coded images processed by corresponding real semantic segmentation images to obtain an optimal weight vector and a bias term of a convolutional neural network classification training model; and in a test stage, inputting an indoor scene image to be semantically segmented into the convolutional neural networkclassification training model to obtain a semantic segmentation prediction graph. According to the invention, the semantic segmentation efficiency and accuracy of the indoor scene image are improved.

An embodiment of the invention discloses an image identification method and device and relates to the technical field of image processing. The method comprises the following steps: carrying out image enhancement on an obtained image to be identified to obtain an enhanced image to be identified; carrying out feature extraction on the enhanced image to be identified according to a preset deep residual neural network to obtain feature information; carrying out Fisher vector coding on the feature information to obtain a Fisher feature vector; and then, identifying the image to be identified according to a preset classifier and the feature vector to obtain an identification result. Therefore, more identification features of the image to be identified can be extracted by utilizing the deep residual neural network and Fisher vector coding; and then, the features are subjected to classification processing, and the image to be identified can be identified quickly according to the classification result. The method is simple to operate, is more efficient and more accurate, and is high in applicability.

The invention discloses a GAN-based medical diagnosis model anti-attack method for solving the security problem of an artificial intelligence medical image diagnosis model. The method comprises the following steps: building a ResNet-101-based high-precision residual neural network diagnosis model for an acquired medical pathological image, and then building a GAN-based confrontation attack network model which comprises a generator G and a discriminator D, wherein the generator G is used for generating a medical image confrontation sample by superposing high-dimensional random noise disturbance x on an input medical image, and the discriminator D is used for discriminating the authenticity of the confrontation sample; employing a PatchGAN discriminator based on a feature extraction image block for designing three layers of feature blocks including a residual block, expansion convolution and a channel attention mechanism as a main method for feature extraction, so that convolution kernel receptive fields of different scales can extract more refined feature map information by using the method and an effective input medical image disturbance area is obtained; therefore, the anti-attack effectiveness of the medical diagnosis model is improved, and the medical diagnosis model can be reinforced and defended from the anti-attack.

The invention provides a memristor network-based residual neural network model and an application method thereof. A memristor network constructed based on a memristor can perform large-scale parallelprocessing, and meanwhile, has huge memory space. The resistance state of the memristor can be flexibly changed by adjusting applied voltage at the two ends of the memristor, and therefore, synaptic plasticity is achieved. The memristor-based memristor network has the advantages of low power consumption, high speed, modularization and the like, and can be built into various neural networks according to the requirements of developers. The memristor network built by the memristor is used as the hardware support of a residual error neural network; on the basis of the memristor network, a residualneural network model is constructed; the functions of training, testing and the like of the residual neural network model are further achieved; and the resistance value of the memristor is controlledby applying voltages to the two sides of the memristor, and therefore, various needed arithmetic operations in the residual neural network model are achieved.

An indoor parking lot remaining parking space prediction method comprises the following steps that 1, acquiring a parking lot unoccupied parking space sequence, and carrying out data preprocessing; step 2, performing denoising processing on the training set of the initial sequence to obtain a time sequence with noise removed, recording the time sequence as a smooth sequence, and dividing the smooth sequence into a test set and a training set; Using the smooth sequence training set to train an LSTM neural network; step 3, constructing and training a gray residual neural network model by using the initial sequence; and step 4, weighting and combining the two prediction models to obtain a final prediction model.

The invention discloses a multi-source domain adaptive model and method based on partial feature alignment, and the method comprises the steps: enabling a feature selection module for partial featureextraction to generate a selection vector of a feature level according to the similarity of each feature dimension of a source domain and a target domain on the basis of a conventional convolutional neural network or residual neural network feature extractor, wherein after the selection vector acts on the initial feature map, part of features highly related to the target domain in the source domain can be screened out. On the basis, the invention further provides three partial feature alignment loss functions for intra-class, inter-domain and inter-class features, so that the purified featuremap has better distinguishability for a classifier, and partial features related to a source domain and a target domain are highlighted. The method is used for the multi-source field self-adaptive classification data set, and compared with an existing multi-source field self-adaptive model, the classification accuracy is higher, and the feature selection effect is better.