475results about How to "Reduce overfitting" patented technology

The invention discloses an automatic insect image identification method based on a depth convolutional neural network. The method comprises the following steps: (1), collecting an original image and carrying out pretreatment to form a training set, and calculating a mean value image of the training set; (2), constructing a depth convolutional neural network; (3), collecting a sub image block randomly from a training sample of the training set and carrying out pre training on the depth convolutional neural network by using the sub image block; (4), training the depth convolutional neural network by using the training set and combining a mini-batch-based random gradient descent algorithm; and (5), carrying out pretreatment on a to-be-measured insect image to form a test sample, and using the trained depth convolutional neural network to identify the test sample after subtracting the mean value image of the training set from the test sample. Therefore, the identification precision is high; the identification types are diversified; the insect within-class change robustness is enhanced; and the insect inter-class similarity sensitivity is high.

The invention discloses a cascaded residual error neural network-based image denoising method. The method comprises the following steps of building a cascaded residual error neural network model, wherein the cascaded residual error neural network model is formed by connecting a plurality of residual error units in series, and each residual error unit comprises a plurality of convolutional layers, active layers after the convolutional layers and unit jump connection units; selecting a training set, and setting training parameters of the cascaded residual error neural network model; training the cascaded residual error neural network model by taking a minimized loss function as a target according to the cascaded residual error neural network model and the training parameters of the cascaded residual error neural 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 cascaded residual error neural 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 gesture recognition method based on 3D-CNN and convolution LSTM. The method comprises the steps that the length of a video input into 3D-CNN is normalized through a time jitter policy; the normalized video is used as input to be fed to 3D-CNN to study the short-term temporal-spatial features of a gesture; based on the short-term temporal-spatial features extracted by 3D-CNN, the long-term temporal-spatial features of the gesture are studied through a two-layer convolutional LSTM network to eliminate the influence of complex backgrounds on gesture recognition; the dimension of the extracted long-term temporal-spatial features are reduced through a spatial pyramid pooling layer (SPP layer), and at the same time the extracted multi-scale features are fed into the full-connection layer of the network; and finally, after a latter multi-modal fusion method, forecast results without the network are averaged and fused to acquire a final forecast score. According to the invention, by learning the temporal-spatial features of the gesture simultaneously, the short-term temporal-spatial features and the long-term temporal-spatial features are combined through different networks; the network is trained through a batch normalization method; and the efficiency and accuracy of gesture recognition are improved.

The invention relates to a dual-channel convolutional neural network (DC-CNN)-based spectral-spatial cooperative classification method for hyperspectral images. For a characteristic that hyperspectral image data adopts a three-dimensional structure, spectral-spatial features are extracted in a mode of combining a one-dimensional convolutional neural network (1D-CNN) channel with a two-dimensional convolutional neural network (2D-CNN) channel to finish the spectral-spatial cooperative classification of the hyperspectral images. For the problem of relatively little artificial mark data of the hyperspectral images, by adopting a data expansion method suitable for the hyperspectral images, the scale of training samples is increased, the training efficiency of a CNN is improved, and the over-fitting problem is reduced.

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 generative adversarial network improved CT medical image pulmonary nodule detection method. The method includes: 1), acquiring a section of a pulmonary CT image; 2), separating according to image morphological properties to acquire a ROI pulmonary parenchyma area; 3), acquiring different suspected pulmonary nodule candidate sets according to a connected domain formed by abinarized image; 4), building a model of an assistant classifier generative adversarial network to generate positive samples overcome the circumstance that positive-negative sample number is unbalanced; 5), building a convolution neural network to classify suspected pulmonary nodule parts to acquire pulmonary nodule areas; 6), using a non-maximum suppression algorithm to acquire a final area of pulmonary nodule. By the method, efficient processing performance of a computer can be fully utilized, certain expandability is provided, and data processing efficiency is improved; through a convolution neural network algorithm, classifying accuracy is improved, CT image data processing performance is improved, and pulmonary nodule images can be built and analyzed more efficiently.

The invention discloses a brain tumor segmentation network and segmentation method based on a U-Net network. The tail of a contraction path of the segmentation network is connected with a spatial pyramid pooling structure; hole convolution of different scales is introduced into a network jump connection part of the segmentation network; an Add operation and original input are adopted to form a residual block with hole convolution; a receptive field of shallow feature information in the contraction path is expanded; fusing with an expansion path of a corresponding stage is carried out. The segmentation method comprises the following steps: cutting and preprocessing a training data set, then constructing a brain tumor segmentation network DCU-Net based on a U-Net network, then inputting a preprocessed two-dimensional image into a segmentation model for feature learning and optimization, obtaining an optimal parameter model of the segmentation model, and finally inputting a to-be-segmented test data set image into the segmentation model for tumor region segmentation. According to the method, the problems of over-segmentation and under-segmentation in brain tumor segmentation can be effectively solved, and the brain tumor segmentation precision is improved.

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

The invention discloses a speaker recognition method based on a convolution neural network and a spectrogram, and the method comprises the following steps: firstly collecting an audio signal of each speaker; secondly converting the audio signals into the spectrogram; thirdly taking an image as an input layer, and training the neural network through AlexNet training; fourthly adjusting the weight values and biases of all layers of the neural network layer by layer through a reverse propagation algorithm; finally obtaining the parameters of the neural network, and classifying the speakers. The method achieves the quick recognition of the speakers through a convolution neural network processing method.

The present invention discloses a feature extraction and state recognition method for one-dimensional physiological signal based on depth learning. The method comprises: establishing a feature extraction and state recognition analysis model DBN of a on-dimensional physiological signal based on depth learning, wherein the DBN model adopts a "pre-training+fine-tuning" training process, and in a pre-training stage, a first RBM is trained firstly and then a well-trained node is used as an input of a second RBM, and then the second RBM is trained, and so forth; and after training of all RBMs is finished, using a BP algorithm to fin-tune a network, and finally inputting an eigenvector output by the DBN into a Softmax classifier, and determining a state of an individual that is incorporated into the one-dimensional physiological signal. The method provided by the present invention effectively solves the problem that in the conventional one-dimensional physiological signal classification process, feature inputs need to be selected manually so that classification precision is low; and through non-linear mapping of the deep confidence network, highly-separable features / feature combinations are automatically obtained for classification, and a better classification effect can be obtained by keeping optimizing the structure of the network.

The invention discloses a text image super-resolution reconstruction method based on a conditional generative adversarial network. According to the method, the structure of a common conditional generative adversarial network is modified, so that the method is more suitable for a super-resolution reconstruction task of a text image; Non-text binary segmentation image is used as extra training supervision information of a super-resolution reconstruction model, and text-text binary segmentation image is combined with extra training supervision information of a super-resolution reconstruction model; the non-text binary segmentation information constructs a loss function of the model to constrain the training of the model, so that the super-resolution reconstruction model is more concentrated in the text part in the image. Compared with a common image super-resolution method, the text image super-resolution reconstruction method disclosed by the invention fully and specifically utilizes theinformation of the text, and the quality of text image super-resolution reconstruction is effectively improved.

The invention discloses an image classification method and an image classification device. The image classification method comprises the following steps: based on a big image data set, training an AlexNet model structure; migrating the five trained convolution layers to a small database to form a lower-level feature extraction layer, and constructing together with a residual network layer, a multiscale pooling layer, a feature layer and a softmax classifier to obtain a migration model structure, wherein the residual network layer includes two convolution layers; inputting small image data set into the migration model structure, upgrading parameters by adopting a batch gradient descending method, and training an image classification hybrid model; and classifying according to the image classification hybrid model to obtain a classification result. By migrating the pretrained convolution layers on the big data set to the small data set, increasing the multiscale pooling layer, and serially connecting the feature quantity output by the residual network layer and the multiscale pooling layer and inputting to the classifier, the feature quantity is increased and the overfitting problem can be relieved; and through the hybrid model trained based on a convolution nerve network and a migration learning, the image classification accuracy can be effectively improved.

The invention relates to a feature selection identification method for non-invasive power load monitoring, which is characterized in that the method comprises the following steps: step 1, detecting aload start-stop event from a total active power of a power consumption scene and extracting a load event feature; step 2, establishing a training data set of a machine learning classify; step 3, carrying out feature selection processing on that train data set, and optimally configuring the features used for the train classification model; step 4, establishing a supervised classification model, andtraining that supervised classification model with the train data set in the step 2 and the feature selected in the step 3; step 5, obtaining that value of each characteristic of the unknown kind ofload event through the step 1, analyzing the machine learn classification model established in the step 4, obtaining the classification result, and finally realizing the detection and identification of the load event. The invention can improve the identification performance of the machine learning classification model established for the identification load event through the feature selection.

The invention discloses a disk fault detection method using multi-model prediction, which extracts multiple characteristics of disk SMART indexes through a sequential data processing means, and establishes a classification model to predict disk state. Step 1, data input: acquiring a data set composed of monitoring data of a plurality of disks in a period of time; step 2, SMART screening: adoptingmutation point detection mode to select SMART index; step 3, feature engineering: using SMART index as the input of the user-defined feature extraction module to extract the features of the SMART index, then extracting the corresponding parameter configuration, and transmitting the parameter configuration to the feature extraction module as a parameter, so as to extract the feature sets of the training set and the test set; step 4, data set balance: desampling the negative sample which occupies a large amount by adopting dimension reduction clustering; step 5, algorithm selection and modeling:on the basis of the step 4, training the classification model and testing whether the current disk belongs to the normal state or the fault state that needs to be replaced.

The invention discloses a high-resolution SAR terrain classification method based on multiscale convolution and feature fusion, and mainly aims at solving the problem in the prior art that the classification precision is low and overfitting easily occurs. The high-resolution SAR terrain classification method comprises the steps of 1, extracting textural features and wavelet features of to-be-classified images; 2, fusing the to-be-classified images, the textural features and the wavelet features to constitute a fusion feature matrix; 3, according to the fusion feature matrix, constructing a training dataset and a testing dataset; 4, adding a multiscale convolution layer and a shuffle layer to an existing CNN network, changing a full-joint layer into a convolution layer, and constructing a multiscale convolution fusion network; 5, using the training dataset to train the multiscale convolution fusion network to obtain model parameters; 6, using the model parameters to initialize the multiscale fusion network to classify a test set. By means of the high-resolution SAR terrain classification method based on the multiscale convolution and the feature fusion, the parameters of the networkare reduced, the overfitting phenomenon of a small sample problem is solved, the classification precision is improved, and the high-resolution SAR terrain classification method can be applied to high-resolution SAR image terrain classification.

The invention discloses a breast cancer cell characteristic analysis system based on deep learning. Based on the deep learning, the system constructs a multilevel convolutional neural network to extract multilevel features, so higher analysis accuracy can be realized; an activation function of a model in the breast cancer cell characteristic analysis system disclosed by the invention is an unsaturated ReLU function, which has faster convergence properties; a pooling layer in the breast cancer cell characteristic analysis system disclosed by the invention adopts an overlapped pooling operation, it can be proved by cross validation that, compared with a traditional non-overlapped pooling layer, the overlapped pooling can further improve the analysis accuracy; and the breast cancer cell characteristic analysis system disclosed by the invention adopts a training mode of sparse autocoder pre-training and Dropout fine tuning to effectively reduce the over fitting of the model and enhance the generalization ability of the trained model, so that the analysis accuracy can be further improved.

The invention belongs to computer visual pattern recognition and a video picture processing method. The computer visual pattern recognition and the video picture processing method comprise the steps that establishing a three-dimensional space-time sub-frame cube in a video human action reorganization model, establishing a human action characteristic space, conducting the clustering processing, updating labels, extracting the three-dimensional space-time sub-frame cube in the video human action reorganization model and the human action reorganization from monitoring video, extracting human action characteristic, confirming category of human sub-action in each video and classifying and merging on videos with sub-category labels. According to the computer visual pattern recognition and the video picture processing method, the highest identification accuracy is improved by 16.5% compared with the current international Hollywood2 human action database. Thus, the video human action reorganization method has the advantages that human action characteristic with higher distinguishing ability, adaptability, universality and invariance property can be extracted automatically, the overfitting phenomenon and the gradient diffusion problem in the neural network are lowered, and the accuracy of human action reorganization in a complex environment is improved effectively; the computer visual pattern recognition and the video picture processing method can be applied to the on-site video surveillance and video content retrieval widely.

The invention relates to a control method in the information processing technical field, in particular to an asynchronous brain-computer interactive control method for an EEG signal. The method comprises the following steps: acquiring EEG signal data under a continuous motion imagery condition; re-sampling training data; extracting features and establishing a classifier model; carrying out on-line feature projection and pattern classification; calculating duration time in a current status according to dynamic change of the duration time by a post-processing technology; and forming closed brain-computer interactive control by asynchronous brain-computer interactive control and neural feedback. The asynchronous brain-computer interactive control method reduces interactive retardance, improves fault-tolerant capability, breaks through a discrete command control mode, and enhances brain-computer interactive flexibility, and can provide a direct brain-computer interactive communication technology for the handicapped or patients with dyskinesia.

The present invention discloses a product quality control method based on integration learning. For prediction of product key quality indexes (yield) in different progresses in a production flow, themethod comprises the following steps of: (1) data analysis based on injection molding data; (2) feature engineering analysis and construction; (3) model design based on integration learning; (4) unbalance processing of data; and (5) multi-model fusion processing scheme. For recommendation of the optimal preset value of process parameters in the production process, the method comprises the following steps of: (6) whole process adjustable parameter recommendation; and (7) for un-adjustable parameters of a special process, recommendation of process adjustable parameters. The product quality control method is suitable for the processing unbalance data in the industrial data to break through a traditional product quality control single parameter analysis mode, and constructs and excavates internal feature relations between parameters through the feature engineering of the machine learning to discover the abnormality of the production process and improve the product quality control.

The invention relates to a multi-peak target tracking method with fusion of background information. According to the method, the background information near the target is utilized, the background information of the same size is extracted from eight concrete positions of upper left, upper, upper right, left, right, lower left, lower and lower right of a current frame target search box to act as candidate target search boxes, the response of the candidate target boxes is computed by using the circulation matrix of the correlation filter tracking method and the Gaussian kernel function, and the classifier parameters are trained and updated so that the success rate in case of rotation change of the target can be enhanced. Existence of the interference situation of similar targets is considered, and the multi-peak target detection strategy, i.e. the possibility that the target is located in any one of multiple peaks of the target output response diagram, is utilized so that the accuracy oftarget tracking can be enhanced.

The invention discloses a pedestrian re-recognition method based on migration learning. The pedestrian re-recognition method includes the steps of 1) preprocessing pictures in a data set by using a multi-scale Retinex color image enhancement algorithm to remove the influence caused by illumination; 2) concentrically extracting relevant features of pedestrians from the preprocessed data set; 3) establishing an asymmetric multi-task discriminant model; 4) refining the model by using a flow-shaped structure without label data; 5 ) optimizing the objective function by using a conjugation gradientmethod to obtain a final classifier; and 6 ) testing the pedestrian re-recognition rate on the classifier based on the pedestrian re-recognition standard database. Few label data and the label-free data are reasonably applied, and the model obtained through training can well improve the pedestrian re-recognition rate.

The invention discloses a vehicle logo locating and identification method fusing a sliding window and a Faster R-CNN. A computer vision technology is used; a vehicle logo is roughly located by identifying a vehicle lamp to cope with the situations that a license plate of a vehicle is shielded and the vehicle does not have the license plate; the vehicle logo identification problem is solved based on the CNN; and positive and negative samples are dynamically generated through RPN, so that the input samples are different, the over-fitting degree of the network is reduced, the robustness of the network is improved, a conventional working mode of identifying the vehicle by the license plate is improved, reliable help is provided for fighting illegal criminal behaviors such as cloned vehicle license plates, one vehicle with multiple license plates, false license plates and the like, the reliability of intelligent traffic is further improved, and a large amount of manpower costs are reduced.

The invention relates to the field of bioinformatics, and discloses a prediction method and system of protein local structure characteristics, which uses the depth learning technology to improve the prediction accuracy, provides key reference information for the prediction of the tertiary structure of a protein, and solves the problem of high cost and low efficiency caused by determining the tertiary structure of the protein by a biological experimental method. The method uniformly constructs characteristic sequences of various protein sequences in a sample set as the input of a training model; the training model adopts a deep neural network model based on stack sparse self-coding with a number of hidden layers of 3, a dropout method is applied to the hidden layers of the whole network, and some neurons in the hidden layers are not allowed to work to reduce the over-fitting of the model; and weight parameters of the training model are optimized through a training set, so that the valueof a constructed loss function is minimized, therefore, the solvent accessibility of each residue or residue contact number prediction in the protein sequences can be carried out correspondingly according to the trained network model.

A fault diagnosis method of a vehicle lithium battery based on a multi-class support vector machine algorithm belongs to the field of vehicle lithium battery fault diagnosis. In order to solve the problem that the existing vehicle lithium battery fault diagnosis technology requires high amount of training data, which leads to the difficulty of realizing real-time on-line fault detection, the method of the invention comprises the following steps of: grouping battery samples into experiments, and sorting the collected data to form a training set and a test set; specifying battery failure criteria; using cross-validation and network search methods to optimize the parameters; constructing a kernel function support vector machine; constructing as biased binary tree five-class support vector machine to get the fault diagnosis model of vehicle lithium battery which can identify different states of lithium battery. The method can quickly and accurately complete the fault diagnosis of a lithiumbattery for a vehicle.

The invention provides a remote sensing image small sample target detection method based on a prototype convolutional neural network. A target detection network mainly comprising a feature extraction and category prototype acquisition module, a prototype guidance RPN module, a redirection feature map module and a detector module is constructed, network model basic learning is performed on a base class data set containing a large number of labeled samples, then a network model is finely adjusted on a balanced sub-data set, and finally, post-processing operations such as non-maximum suppression and the like are performed to realize multi-class target detection of the small sample remote sensing image. According to the method, different types of targets can be rapidly and accurately detected from the optical remote sensing image with the complex background by using a small amount of new class annotation data, and the method has relatively high detection precision and relatively high detection speed.

The invention discloses an encrypted anonymous network traffic identification method. The method comprises the following steps: 1) extracting multi-granularity level characteristics of each flow filefrom an encrypted anonymous network flow data set, wherein the multi-granularity level characteristics comprise flow characteristics, packet characteristics, host behavior characteristics, TCP headerrelated characteristics and IP header related characteristics; 2) filtering the features obtained in the step (1), and filtering redundant features and features irrelevant to flow identification or features with the relevancy lower than a set threshold value; and 3) training an XGBoost model by utilizing the characteristics selected in the step 2), and then identifying the anonymous network traffic to be identified by utilizing the XGBoost model. The method is superior to the existing baseline identification method in the aspects of overall accuracy, precision rate, recall rate and F1 value onmodel performance.

The invention discloses a method for reducing dimensions of hyper-spectral data on the basis of pairwise constraint discriminate analysis and non-negative sparse divergence, and belongs to methods for processing hyper-spectral remote sensing images. The method aims to solve the problem of deterioration of the classification performance of most advanced algorithms for classifying hyper-spectral data on the basis of machine learning when source hyper-spectral data and target hyper-spectral data are distributed differently. The method includes firstly, performing pairwise constraint discriminate analysis according to pairwise constraint samples; secondly, designing a non-negative sparse divergence criterion to create a bridge among source-field hyper-spectral data and target-field hyper-spectral data which are distributed differently; thirdly, combining the pairwise constraint discriminate analysis with the bridge to transfer knowledge from the source hyper-spectral data to the target hyper-spectral data. The pairwise constraint samples containing discriminate information can be automatically acquired. The method has the advantages that the knowledge can be transferred among the hyper-spectral data acquired at different moments, in different areas or by different sensors; the information of the source-field hyper-spectral data can be effectively utilized to analyze the target-field hyper-spectral data, and high integral classification precision and a high Kappa coefficient can be acquired.

The invention discloses a financial reimbursement all-invoice picture recognition and processing method. According to the method, a specific scene of invoice recognition is optimized to some extent, invoice recognition types are expanded, invoices of all types can be recognized, and recognition is more accurate and higher in efficiency. The method is mainly used for performing recognition and processing after the invoices of all types are scanned into pictures. The method comprises the specific steps that the invoice pictures obtained after scanning are subjected to color preprocessing, and input data is provided for picture contour preprocessing; the pictures obtained after color preprocessing are subjected to contour detection, and pixel interference outside an invoice paper range is eliminated; the pictures obtained after contour detection are subjected to text preprocessing, characters on the invoices are recognized; and a character set is formed and output; the output character set is subjected to picture recognition and processing, a recognition and processing result is organized into a result object with semantics according to semantics of invoice recognition, and the result object is used as a final recognition result.

The invention discloses a pedestrian re-identification and method for a binary triple twin network model, and the method comprises the steps: enabling three paths of convolutional neural networks to input positive and negative samples and detection sample images, extracting image features, and enabling each path of convolutional neural network to comprise a convolutional layer, a pooling layer, and a full connection layer; wherein the convolutional layer, the pooling layer and the full connection layer of each convolutional neural network are the same and share weight parameters, and binarizing the weight parameters and an activation function value between the convolutional layer and the pooling layer; connecting the Softmax layer with each full connection layer, and classifying and normalizying features output by the convolutional neural network and connecting the Triplet loss verification function module with the Softmax layer and using the module for receiving the sample characteristics output by the normalization classification layer and carrying out similarity calculation on the sample pairs. According to the invention, the deep learning model with stronger identification capability is obtained, and the problem of large content difference caused by illumination, scene change and human body posture diversification of similar and same pedestrian pictures on different pedestrian picture contents is better solved.

The invention discloses a remote sensing image classification method, a storage medium and computing equipment, and the method comprises the steps: building a remote sensing image set, carrying out the standardization of the remote sensing image set, and obtaining a training sample set and a test sample set; setting a multi-scale feature extraction module, and generating feature maps of two scalesby setting different hole convolution in two parallel convolution modules; setting a self-adaptive feature fusion module, wherein the self-adaptive feature fusion module can adaptively select usefulinformation in the two generated features of different scales and perform fusion; building a whole neural network model; performing iterative training on the whole neural network model by using the training sample set; and randomly selecting a sample from the test samples as a position category sample, and classifying unknown samples needing to be predicted by using the trained neural network. According to the method, redundant information is reduced, multi-scale features are selected more flexibly, the stability of the network is improved, and then the classification capacity of the network model is improved.