121results about How to "Avoid Gradient Explosion" patented technology

The invention discloses a fundus image retinal vessel segmentation method and a fundus image retinal vessel segmentation system based on deep learning. The fundus image retinal vessel segmentation method comprises the steps of performing data amplification on a training set, enhancing an image, training a convolutional neural network by using the training set, segmenting the image by using a convolutional neural network segmentation model to obtain a segmentation result, training a random forest classifier by using features of the convolutional neural network, extracting a last layer of convolutional layer output from the convolutional neural network, using the convolutional layer output as input of the random forest classifier for pixel classification to obtain another segmentation result, and fusing the two segmentation results to obtain a final segmentation image. Compared with the traditional vessel segmentation method, the fundus image retinal vessel segmentation method uses the deep convolutional neural network for feature extraction, the extracted features are more sufficient, and the segmentation precision and efficiency are higher.

The invention discloses an ReLU convolutional neutral network-based image denoising method. The method comprises the following steps of building an ReLU convolutional neutral network model, wherein the ReLU convolutional neutral network model comprises a plurality of convolutional layers and active layers after the convolutional layers, wherein the active layers are ReLU functions; selecting a training set, and setting training parameters of the ReLU convolutional neutral network model; training the ReLU convolutional neutral network model by taking a minimized loss function as a target according to the ReLU convolutional neutral network model and the training parameters of the ReLU convolutional neutral network model to form an image denoising neural network model; and inputting a to-be-processed image to the image denoising neural network model, and outputting a denoised image. According to the ReLU convolutional neutral network-based image denoising method disclosed by the invention, the learning ability of the neural network is greatly enhanced, accurate mapping from noisy images to clean images is established, and real-time denoising can be realized.

The invention discloses a cutter wear state monitoring method based on a deep gated cycle unit neural network. The method comprises the steps that vibration signals generated in the tool machining process are collected in real time through a sensor, after wavelet threshold denoising, the signals are input into a one-dimensional convolutional neural network for single time step time sequence signallocal feature extraction; then, inputting the time series signal into an improved deep gated recurrent unit neural network CABGRUs to carry out time series signal time series feature extraction; an Attention mechanism is introduced to calculate network weights and reasonably distribute the network weights, and finally, signal feature information with different weights is put into a Softmax classifier to classify tool wear states, so that complexity and limitation caused by manual feature extraction are avoided; meanwhile, the problem that a single convolutional neural network ignores correlation before and after a time sequence signal is effectively solved, and the accuracy of the model is improved by introducing an Attention mechanism. Therefore, the method has the characteristic of improving the real-time performance and accuracy of cutter wear state monitoring.

The invention discloses a two-dimensional recursive network-based recognition method of Chinese text in natural scene images. Firstly, a training sample set is acquired, and a neural network formed bysequentially connecting a deep convolutional network, a two-dimensional recursive network used for encoding, a two-dimensional recursive network used for decoding and a CTC model is trained; test samples are input into the trained deep convolutional network, and feature maps of the test samples are acquired; the feature maps of the test samples are input into the trained two-dimensional recursivenetwork, which is used for encoding, to obtain encoding feature maps of the test samples; the encoding feature maps of the test samples are input into the trained two-dimensional recursive network, which is used for decoding, to obtain a probability result of each commonly used Chinese character in each image of the test samples; and clustering searching processing is carried out, and finally, the overall Chinese text in the test samples is recognized. According to the method of the invention, space / time information and context information of the text images are fully utilized, the text imagepre-segmentation problem can be avoided, and recognition accuracy is improved.

The invention relates to a time-space domain correlation prediction method for air pollutant concentration, which comprises the steps of S1, constructing a prediction model based on a residual error network and a convolutional LSTM network by taking PM2.5 as a sample for target pollutant prediction; s2, selecting appropriate training and testing data from the environment monitoring data to complete initialization of the prediction model; s3, training the prediction model stage by stage to obtain a neural network prediction model capable of accurately predicting PM2.5; s4, selecting hyper-parameters (the number of layers, the number of nodes and the learning rate) of the model by utilizing the verification set until the model is optimal; and S5, carrying out urban PM2.5 prediction by utilizing the verified prediction model. Compared with the prior art, the method has the advantages that the convolutional LSTM network is used as a middle layer, deep space-time association feature extraction is performed on spatial features extracted by the bottom ResNet network, accordingly, the prediction performance of the network model can be improved, the hidden state of the convolutional LSTM can be received by the aid of the full connection layer, and a final prediction result can be generated.

The invention discloses a text classification method based on a capsule network. The problems that in the prior art, the overall precision is not high, the applicability is not high, a large amount ofimportant information is lost in the feature extraction process, and the relation between the local part and the overall part in a text is ignored are solved. The method comprises the following steps: 1, nodes in a capsule network being capsules consisting of a group of neurons, executing complex internal calculation on input by using matrix capsules, outputting instantiation parameters from results in a matrix form, and meanwhile, outputting an activation value of each capsule is output; 2, calculating between two adjacent layers in the capsule network through an EM routing algorithm; representing a higher-dimensional concept through Gaussian cluster, and each activation capsule selecting a capsule of the next layer as a father node through an iterative routing process, so that link prediction is realized between two adjacent layers of networks; and 3, training a weight parameter of the full connection layer, and calculating the prediction probability of the real label by using a softmax activation function.

The invention belongs to the field of hyper-spectral intelligent perception, and particularly discloses a spatial spectrum fusion hyper-spectral image classification method based on a three-dimensional deep residual network, which comprises the following steps: S1, generating candidate frames by using a sliding window method, and generating a plurality of windows; S2, randomly dividing the windowinto a training set and test set data; S3, training a three-dimensional depth residual network (3D-CNN) based on the hyperspectral data in the training set; S4, inputting a test set sample into the classification model, The hyperspectral image classification method has the advantages that the characteristics of the input data are extracted and predicted, the spectral characteristics and the spatial spectrum characteristics of the hyperspectral image are extracted at the same time, the classification precision of the hyperspectral image is further improved, a residual network structure is introduced, and the problem of learning degradation in a traditional hyperspectral classification neural network is solved. The hyperspectral image target classification method is clear in structure and easy to implement, the structural characteristics of the hyperspectral image can be fully utilized, and the hyperspectral image target classification precision is remarkably improved while the calculation time is shortened.

The invention relates to the field of biological recognition, trains a face recognition model based on deep learning, and particularly discloses a face recognition model training method and device, aface recognition method and device, equipment and a storage medium. The face recognition model training method comprises the steps: carrying out the training of a preset convolutional neural network,so as to construct a feature extraction network; establishing connection between the feature extraction network and a preset classification network to obtain a first convolutional neural network model; freezing a weight parameter of the feature extraction network of the first convolutional neural network model; performing iterative training on a classification network in the first convolutional neural network model to obtain a second convolutional neural network model; unfreezing the weight parameters of the feature extraction network of the second convolutional neural network model; and training the unfrozen second convolutional neural network model to obtain a face recognition model. The face recognition model training method can improve the face recognition speed and improve the stability of the model.

The invention discloses a macular location method based on improved Faster R-CNN, which includes collecting training samples, constructing a network model, training the network model, constructing a detection model and performing macular detection and location. The invention utilizes the improved Faster R-CNNto achieve effective localization of macular region, reduces the influence of optic disc and blood vessel on macular region and has strong anti-noise ability, which greatly improves the accurate localization of macular region and lays a foundation for the subsequent analysis and processingof fundus images.

The invention discloses an aspect-level text sentiment classification method and system, and the method comprises the steps: extracting the long-distance dependence features of a sentence text according to the obtained local feature vectors of the sentence text, and obtaining the context feature representation of the sentence text; constructing a syntactic dependency relationship among words in the sentence text according to the context feature representation of the sentence text to obtain aspect-level feature representation of the sentence text; and constructing a dependency tree-based graphattention neural network, and obtaining aspect-level emotion categories of the text according to aspect-level feature representation of the sentence text. The method comprises the steps of extractinglocal feature information in a sentence by adopting a convolutional neural network, learning pooled features of the convolutional neural network by utilizing a bidirectional long-short-term memory network, obtaining context information of the sentence, constructing a dependency tree-based graph attention network model, and modeling a sentence dependency relationship by utilizing syntactic information of a dependency tree, thereby improving the performance of sentiment classification.

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 relates to a navigation reminding method for short-term traffic flow prediction based on SVD-PSO-LSTM. Firstly, an LSTM model is trained and optimized; collecting historical traffic flowdata; preprocessing the LSTM model, and inputting the preprocessed LSTM model into the trained and optimized LSTM model; the method comprises the following steps: outputting a prediction result of ashort-time traffic flow by a traffic flow sensor, publishing the prediction result of the short-time traffic flow on navigation software, displaying the prediction result of the short-time traffic flow on a road section in different colors according to the traffic flow, providing displayed information for a driver to refer to congestion, and reasonably planning travel and selecting a navigation route according to the displayed information. When the LSTM model is trained and optimized, historical traffic flow data is collected, preprocessed and then divided into a training set and a test set, and then the LSTM model is trained and optimized by adopting a PSO algorithm and the training set; wherein the preprocessing comprises the step of performing noise reduction operation on the data by adopting an SVD algorithm. The method provided by the invention is more accurate in prediction effect, and can reasonably plan a navigation route.

The invention provides a hospitalization behavior prediction method and device based on time-varying attention improved Bi-LSTM. The hospitalization behavior prediction method comprises the steps of extracting features from massive medical insurance data, wherein the correlation degree between the features and hospitalization behaviors is greater than a preset correlation degree threshold; constructing Bi-LSTM by utilizing the extracted hospitalization behavior characteristics and the corresponding weights of the hospitalization behavior characteristics; updating a weight value of each hospitalization state prediction data in the Bi-LSTM by adopting hospital-disease attraction data pre-generated by an attention mechanism based on the hospitalization state prediction data obtained by the Bi-LSTM; constructing a time adjustment function; outputting a multi-period multi-state hospitalization state prediction vector; utilizing the hospitalization state prediction vector to construct a softmax prediction function; calculating a loss function of an output value of the softmax prediction function, training learning parameters of Bi-LSTM by adopting back propagation, and completing training of the model; and after model training is completed, outputting a prediction result of the experimental sample set, comparing the prediction result with an actual hospitalization behavior, and feeding back and updating a weight value of hospitalization state prediction data.

The invention provides a thermal power plant condenser vacuum degree prediction method based on multilayer LSTM. Firstly, a sample data set is constructed, input data and output data in a sample are subjected to standardization processing respectively, then the input data sequentially enter a multi-layer long-short-term memory neural network, and model training is performed by using an adaptive moment estimation algorithm; and finally, to-be-predicted data is input into the trained prediction model, and the vacuum degree of the condenser is predicted. According to the method, the long-term andshort-term memory neural network structure is applied to condenser vacuum degree analysis, deep mining of big data is achieved, the accuracy and speed of data prediction are improved, and use of system resources is optimized. According to the two-layer LSTM structure designed by the invention, the model depth can be increased, the model capacity can be increased, the prediction precision can be improved, and meanwhile, the adaptive moment estimation algorithm is used for optimizing the traditional gradient descent algorithm, so that the problem of gradient explosion is avoided.

The present invention discloses a Chinese text overall recognition method in a natural scene image. The method comprises: acquiring a training sample set, and training a neutral network formed by sequentially connecting a deep convolution network, a multi-layer bidirectional recursive network for encoding, a multi-layer bidirectional recursive network for decoding, and a CTC model; inputting a test sample into the trained deep convolution network to obtain each feature sequence frame of the test sample; inputting each feature sequence frame of the test sample into the trained multi-layer bidirectional recursive network for encoding to obtain each code sequence frame of the test sample; inputting each code sequence frame of the test sample to the trained multi-layer bidirectional recursivenetwork for decoding, and obtaining each common Chinese character in each frame image of the test sample as a probability result of the Chinese character in the frame image; and carrying out cluster search processing to finally recognize the overall Chinese text in the test sample. According to the method disclosed by the present invention, the accuracy of the Chinese text overall recognition in the natural scene image is improved.

The invention discloses a method for recognizing lithology through reconstructed textures. The method comprises the following steps: establishing an original specimen image library; establishing a standardized image library; extracting features of the rock specimen image by using discrete convolution and maximum pooling operation of a convolutional neural network MobileNet-V2 convolutional model,and effectively preventing the problems of gradient explosion and gradient disappearance in the image feature extraction process by using a new algorithm; and selecting the rock category with the maximum probability to output a result, and completing the recognition process. According to the method, the problems of low precision and low recognition efficiency caused by the fact that an existing algorithm is limited to two dimensions in the field of randomly recognizing rock lithology under the conditions of particle distribution and section texture development and imaging quality including rock color, definition and the like are solved. The method has the advantages that various lithology characters can be widely identified without being limited by external factors such as backgrounds andcolors, and the identification accuracy is improved; intelligent identification operation speed improvement.

The invention discloses an unreal information detection method based on a BERT model and an enhanced hybrid neural network. The method comprises the steps of preprocessing a to-be-detected text; performing convolution and pooling operation on the input matrix by using a CNN network, and splicing the input matrix into a feature sequence; taking the feature sequence as the input of a BiLSTM network,and comprehensively capturing the deep semantic features of the text from the front direction and the rear direction by using a forward LSTM unit and a backward LSTM unit respectively; generating a semantic code containing attention distribution by utilizing an attention layer, and optimizing a feature vector; and finishing classification detection of the feature vectors by utilizing a classifierof the output layer, and judging whether the feature vectors are the non-real information. According to the method, the CNN, the BiLSTM and the attention mechanism are combined, the detection precision of the unreal information is high, local phrase features and global context features of the text of the unreal information can be extracted, text keywords can also be extracted, and the unreasonable influence of irrelevant information on a detection result is reduced.

The invention discloses a many-to-many voice conversion method and system based on speaker style feature modeling. Firstly, a multi-layer sensor and a style encoder are added to a StarGAN neural network, effective extraction and constraint of speaker style features are realized, the defect of limited speaker information carried by one-hot vectors in a traditional model is overcome, then, an adaptive instance normalization method is adopted to realize full fusion of semantic features and speaker personality features, so that the network can learn more semantic information and speaker personality information, furthermore, a lightweight network module SKNet is introduced into the generator residual error network, so that the network can adaptively adjust the size of a receptive field according to multiple scales of input information, the weight of each feature channel is adjusted through an attention mechanism, the learning ability of frequency spectrum features is enhanced, and the details of the frequency spectrum features are refined.

The invention discloses a road scene semantic segmentation method based on a full-residual cavity convolutional neural network. The method comprises the steps: constructing the full-residual cavity convolutional neural network at a training stage, and enabling the full-residual cavity convolutional neural network to comprise an input layer, a hidden layer and an output layer, and enabling the hidden layer to comprise one transition convolution block, eight neural network blocks, seven deconvolution blocks and four fusion layers; inputting each original road scene image in the training set intoa full residual cavity convolutional neural network for training to obtain 12 semantic segmentation prediction images corresponding to each original road scene image; calculating a loss function value between a set formed by 12 semantic segmentation prediction images corresponding to each original road scene image and a set formed by 12 one-hot coded images processed by corresponding real semantic segmentation images to obtain a full residual cavity convolutional neural network training model; in the test stage, a full residual cavity convolutional neural network training model is used for prediction; the method has the advantages of high segmentation accuracy and strong robustness.

The invention relates to a wind driven generator fault diagnosis method based on a long-term and short-term memory model recurrent neural network. The method comprises the following steps of: 1, carrying out the modeling of a fan reference system, enabling the fan reference system to be subdivided into eight types of faults, and setting the fault types and occurrence time, 2, simulating the fan reference model to obtain an actual measurement value, 3, performing data preprocessing on the actual measurement value obtained in the step 2, and constructing a data sample set of the actual measurement value, and 4, building an LSTM model, training the preprocessed data in the step 3 by using the LSTM model, continuously adjusting parameters in a training process, evaluating a training effect byusing an equipartition error, comparing a predicted value with the actual measurement value of the sensor obtained in the step 2, and setting a threshold value to judge the fault occurrence time and position.

The invention discloses an ancient poem automatic identification method based on deep learning, and the method is characterized in that the method comprises the steps:by including collecting trainingcorpora, carrying out the data preprocessing, carrying out the embedding of feature vectors, carrying out the training of a neural network, and carrying out the automatic identification of ancient poems. The deep neural network model in the text classification form is constructed, whether the text sentences are ancient poetry types or not is automatically recognized, and meanwhile it can be effectively avoided that wrongly written characters are wrong, and the recognition accuracy is reduced. And the requirements of application scenes such as poetry quality detection, literature work classification management and automatic collection of ancient poetries on an ancient poetry automatic identification technology can be met.

The invention discloses a cutter wear state real-time monitoring method, which comprises the following steps: acquiring an original vibration signal in real time by using an acceleration sensor, and continuously sampling and cutting the original vibration signal into a time sequence signal consisting of 2000 sampling points (2000, 3); taking the time sequence signal with the size of (2000, 3) as input data, inputting the input data into a one-dimensional convolutional neural network (CNN) for neighborhood filtering, performing calculation by using a sliding window, and finally obtaining a high-dimensional feature of a single time step time sequence signal; processing high-dimensional features generated by continuous time step time sequence signals by adopting an improved BiLSTM (Bipolar Long Short Term Memory) network; introducing an Attention mechanism to calculate importance distribution of continuous time step time sequence signal features, and generating a time sequence signal feature model containing attention probability distribution; and training the network model to obtain a wear classification result. The method has the characteristic of improving the identification precision and generalization performance of real-time monitoring of the network model.

The invention discloses a multi-source remote sensing image pixel-by-pixel classification method based on a correlation fusion network, and the method comprises the steps: reading a multi-spectral image from a data set,wherein the multi-spectral image comprises registered PAN and MS image data and a corresponding class sign true image with only a part of regions; marking an edge sample; respectively preprocessing the PAN image and the MS image, and selecting a training set and a test set; constructing a fusion network model of multi-source remote sensing image pixel-by-pixel classification; constructing a network loss function; training the classification model by using the training data set to obtain a trained classification model; and classifying the test data set by using the trained classification model to obtain the category of each pixel point in the test data set. According to the invention, the classification performance is improved, and the improved loss function strategy improves the pixel-by-pixel classification performance of the remote sensing image.

The invention discloses a many-to-many speaker conversion method based on Transitive STARGAN. According to the invention, a STARGAN generator is combined with a transmission network, features extracted by a coding network are transmitted to a corresponding network layer of a decoding network in the generator; the learning capability of the decoding network on semantic features of different scalesis improved; the learning function of the model on the frequency spectrum deep features is realized; the frequency spectrum generation quality of the decoding network is improved, semantic features and personalized features of speakers are learned more fully, so that the personality similarity and voice quality of converted and synthesized voice are better improved, the problem that the personality similarity and the naturalness are poor after STARGAN model conversion is solved, and high-quality many-to-many speaker conversion under a non-parallel text condition is realized.

The invention discloses a code method-level software defect positioning method, which belongs to the field of software maintenance, and comprises the following steps of: constructing and training a seq2seq model; constructing a defect positioning model based on the seq2seq model; constructing a training sample of the defect positioning model; training a defect positioning model by using the training sample; and predicting the defect method in the defect report by utilizing the trained defect positioning model. According to the method provided by the invention, historical data and a deep learning technology are combined to train a defect positioning model; according to the method, a seq2seq model is utilized to determine parameters of a code encoder in the defect positioning model, code method annotation is used as supervision information, the information amount is rich, the purpose of positioning software defects on the method level is well achieved, and the software defect repairing efficiency can be greatly improved.

The invention discloses a road intersection steering ratio prediction method based on an LSTM neural network. The road intersection steering ratio prediction method comprises the steps: firstly carrying out the statistics of traffic flows in different flow directions of all entrance lanes of a road intersection, and carrying out the data preprocessing; then, calculating the steering ratios of allthe entrance lanes of the road intersection in all the flow directions, and splitting a training set and a test set; then, designing a prediction model based on an LSTM neural network, and selecting an appropriate activation function and an appropriate loss function; and finally, sequentially selecting the training set and the test set of different entrance lanes, training and testing the corresponding prediction model, and completing the establishment of the prediction model, thereby being applied to the prediction of the road intersection steering ratio. According to the road intersection steering ratio prediction method, the time sequence characteristic of the traffic flow can be fully utilized, and the prediction precision of the road intersection steering ratio is effectively improved, and meanwhile, the problems of gradient explosion and gradient dispersion are avoided through the reasonable neural network structure and the good activation function characteristic, and the applicability of the prediction method is improved.

The invention discloses an SAR image ship target detection method and system under a passive interference scene. The method comprises the steps: dividing an SAR image ship target detection data set under a passive interference condition into a training sample and a test sample; constructing a Faster R-CNN detection model with attention enhancement; constructing an overall loss function of a region-of-interest classification task, a region-of-interest position regression task, a target classification task, a target position regression task and a target region saliency detection task of the attention-enhanced Faster R-CNN detection model; training the attention-enhanced Faster R-CNN detection model added with the overall loss function L by using the training sample; inputting the test sample into the attention-enhanced Faster R-CNN detection model containing the weight parameters after training, predicting and outputting a target bounding box position, a target category and a confidence score of the target in the test sample, and completing SAR image ship target detection. According to the method, the SAR image ship target detection precision in the passive interference scene is improved.

The invention belongs to the field of computer vision, and relates to a video behavior identification method based on a compression reward and punishment mechanism. The technical problems that an existing video behavior recognition method is large in calculated amount, poor in robustness, low in accuracy and the like are mainly solved. According to the invention, a convolutional neural network containing a compression reward and punishment mechanism is designed for video behavior identification. The network is constructed based on a time segmentation network. The method comprises the followingsteps: firstly, dividing a video into three segments, randomly extracting an optical flow image and an RGB frame from each segment, respectively inputting the optical flow image and the RGB frame into a time and space network, weighting the extracted features through compression and reward and punishment operations, and respectively carrying out preliminary prediction on behaviors of the weightedtime and space features on a time channel and a space channel; fusing the preliminary prediction result of each segment to obtain a video-level prediction result; and finally, fusing the video-levelprediction results to obtain a video behavior recognition result. Experiments are carried out on data sets UCF101 and HMDB51, and results show that compared with other models, the model has higher accuracy.

The invention discloses an image super-resolution reconstruction method and system based on a convolutional neural network. The method comprises the steps of obtaining a to-be-reconstructed image; preprocessing the to-be-reconstructed image to obtain a preprocessed image; extracting features of the preprocessed image to obtain a feature image set; performing nonlinear mapping on the feature imageset to obtain a global feature mapping set; and performing reconstruction according to the global feature mapping set to obtain a reconstructed high-resolution image. By adopting the method or the system, the reconstruction quality of the image can be improved.

The invention discloses a tail end space energy consumption prediction method based on building total energy consumption, a medium and equipment, and the method comprises the steps of sequentially transmitting preprocessed sample data of N tail end spaces at a moment t and first tau time steps into a tail end space energy consumption prediction model in each step of model training; enabling the model to obtain N tail end space energy consumption prediction values at the t moment through N times of forward calculation, adding the N tail end space energy consumption prediction values to the actual total energy consumption of the building to calculate a loss function, and adjusting parameters of the tail end space energy consumption prediction model through back propagation of a gradient descent method; repeating the training process on the sample data at all moments until the model converges to the prediction precision, and completing the training of the end space energy consumption prediction model; and predicting the energy consumption generated by the tail end space by using the obtained tail end space energy consumption prediction model through the controller parameters of the tail end devices, the temperature and humidity sensors, the Internet weather information, the people flow density, the power of the electric appliances and the lighting devices and the house structure parameters in the operation process of the building heating and ventilation system. According to the invention, the training of the tail end space energy consumption prediction model is realized underthe condition that the tail end space energy consumption historical data is missing.