177results about How to "Prevent overfitting" patented technology

The invention discloses an unbalanced data classification method based on unbalanced classification indexes and integrated learning, and mainly solves the problem of low classification accuracy of the minority class of the unbalanced data in the prior art. The method comprises steps as follows: (1), a training set and a testing set are selected; (2), training sample weight is initialized; (3), part of training samples is selected according to the training sample weight for training a weak classifier, and the well trained weak classifier is used for classifying all training samples; (4), the classification error rate of the weak classifier on the training set is calculated, is compared with a set threshold value and is optimized; (5), voting weight of the weak classifier is calculated according to the error rate, and the training sample weight is updated; (6), whether the training of the weak classifier reaches the maximum number of iterations is judged, if the training of the weak classifier reaches the maximum number of iterations, a strong classifier is calculated according to the weak classifier and the voting weight of the weak classifier, and otherwise, the operation returns to the step (3). The classification accuracy of the minority class is improved, and the method can be applied to classification of the unbalanced data.

The invention particularly relates to a sentiment dictionary learning based text sentiment analysis method and system. The method includes steps: subjecting initial text data to normalization processing to generate preprocessed text data, and clustering the preprocessed text data to a preset field; adopting a seed sentiment dictionary based sentiment word discovery method to form a special sentiment dictionary in the preset field; subjecting the preprocessed text data to retrieval according to the special sentiment dictionary to acquire target text data serving as initial training corpus in the corresponding preset field, and classifying inputted to-be-classified texts through formed multiple classifiers. Labor cost is reduced, the problem of overfitting caused by a single classifier is avoided, and accuracy in text sentiment analysis is improved by consideration of the text related fields.

The invention discloses a method and system for detecting and locating network anomaly, which relates to the fields of Internet security, deep learning and neural network. The method comprises the following steps: firstly, dividing the URL according to special characters; secondly, performing word vector encoding on the divided URL with word2vec; putting word vectors into a convolution layer for automatic feature processing; combining a convolution layer result with an attention layer which possesses a sequential attention mechanism; and finally, performing maximum pooling and full-connectionon an attention layer result to obtain a final anomaly detection result, and at the same time, using the output of the attention layer to locate the malicious code in the URL. The invention has an excellent detection effect, not only is the detection rate high, but also the malicious code fragment in the URL can be located and visualized, thus effectively avoiding the drawbacks of the artificial feature engineering and the expert knowledge method.

The invention discloses a far infrared pedestrian detection method for changed scenes. A sample extension target data set is screened out of auxiliary data on the basis of the Boosting-style inductive transfer learning algorithm DTL Boost. At first, a sample importance measurement model based on the k-nearest neighbor is utilized for evaluating the similarity between the auxiliary data and the target data, and corresponding initial weights are distributed for different samples in the auxiliary data. In the training process, the prediction inconsistency degree of member classifiers is explicitly defined, iterative updating is carried out on the current weights of the auxiliary data and the target data sample through the prediction error rate of the current member classifiers, a sample extension training set with the positive transfer ability is screened out of the auxiliary data, and the different member classifiers are encouraged to learn different parts or aspects of the target data. In this way, an integrated classifier with the stronger generalization ability is obtained through training, and the robustness of pedestrian detection in the new scene is enhanced.

The invention belongs to the technical field of communication control and communication processing, and discloses an isolated forest-based binary classification abnormal point detection method and aninformation data processing terminal. The method comprises the steps of carrying out initial static average blocking on an original data set, and calculating the density in the block and the mean density; after calculating the density in each block of the static block, reducing the data set by taking the mean density of the original data set as a threshold value; constructing an isolated forest byusing a node recursion method; performing corresponding feature extraction and datamation on the original data set, and calculating the spatial position distances between the clustering center pointand other points; adding the abnormal score calculated on the basis of the density and the distance and the abnormal score calculated on the basis of the proof information and comparing with a corresponding threshold value. According to the method, the accuracy of an abnormal point detection algorithm is effectively improved, the actual data size in the abnormal detection process can be greatly reduced, the calculation resources are saved, and the abnormal detection efficiency is improved, and the robustness of an abnormal detection algorithm is enhanced.

The invention discloses a computer aided model construction method based on deep learning gastric cancer pathological sections, and belongs to the technical field of artificial intelligence. The method uses a 121-layer dense-connected convolutional neural network to perform image recognition. A dense block structure in a DenseNet allows the high-level part of the network to acquire shallow features, which greatly reduces the over-fitting phenomenon. At the same time, the model has a large number of layers, which can fit more complex and smoother decision functions. Although the number of layers is large, the number of parameters of the model is not large, which saves resource consumption. In order to further avoid over-fitting, a training mechanism for migration learning is adopted. The model will be pre-trained on an ImageNet dataset to give the model a strong image feature extraction capability. The main optimization of the model during formal training can be better focused on how toextract the features of the diseased area, and the utilization rate of the data is greatly improved.

The invention discloses a human face identification method and apparatus, and belongs to the field of human face identification. The method comprises: performing feature extraction on a to-be-identified human face image by using a plurality of pre-trained convolutional neural networks to obtain a plurality of sub-feature vectors of the to-be-identified human face image, wherein the sub-feature vectors of the to-be-identified human face image are same in number of dimensions; normalizing the sub-feature vectors of the to-be-identified human face image; performing addition on the normalized sub-feature vectors of the to-be-identified human face image, and multiplying the sum of the normalized sub-feature vectors by a coefficient to obtain a union feature vector of the to-be-identified human face image; and performing human face identification by using the union feature vector of the to-be-identified human face image or/and the sub-feature vectors of the to-be-identified human face image. According to the human face identification method and apparatus, the training time of the convolutional neural networks is shortened, the over-fitting of the convolutional neural networks is avoided, and the operation is simple and convenient; and identification modes are more diversified and the accuracy is higher.

The invention discloses a rotating machine fault diagnosis method based on an optimized structure convolution neural network. Firstly, working signals are collected in the normal state and the fault state of a rotating machine, then are converted into gray scale images, and the gray scale images and the corresponding fault labels are used as training samples to train the constructed convolution neural network; and in the working process of the rotating machine, a working signal is collected and is converted into a gray scale image, and the trained convolution neural network is input for faultdiagnosis. The invention adopts the method of converting the collected working signals of the rotating machine into gray-scale images, and completes the task of multi-classification of faults of the rotating machine through a convolution neural network.

The invention discloses a pedestrian re-identification model, method and system for adaptive difficulty mining. The identification method comprises the steps of: randomly dividing sample pictures intoa training set used for each iteration, inputting the training set into a convolutional neural network, obtaining the probability that each sample pair belongs to a positive or negative sample pair by using a softmax function, and then obtaining the loss of each sample pair by using a multinomial logistic function; obtaining a difficult sample pair by using the loss of each sample pair; and training the convolutional neural network by using the difficult sample pair until the current number of iterations reaches the upper limit of the number of iterations, thus obtaining the pedestrian re-identification model. The pedestrian re-identification model is used to extract features of each picture in a picture set to be identified, and then a similarity order of the sample pairs in the pictureset to be identified is obtained. The pedestrian re-identification model, method and system avoid over-fitting and under-fitting, and have high recognition accuracy.

The invention discloses a migration convolution neural network method for face expression recognition, which comprises the following steps: S1, acquiring a face expression image data set, dividing thedata set into a training set, a verification set and a test set; 2, cascade that migration network and the convolution neural network to construct a migration convolution neural network model; The input of MCN model is facial expression image data, and the output is facial expression category. 3, train that migration convolution neural network model by using a training set, and optimize the trained migration convolution neural network model by using a verification set; S4, the facial expression recognition accuracy of the test set is tested by the optimized migration convolution neural network model. The invention utilizes the migration convolution neural network to recognize the facial expression, and solves the problems that the small data set cannot be converged and over-fitted on a large number of networks.

The invention discloses a bottom material sonar image classification method based on a convolution neural network, belonging to the technical field of image classification. The method comprises obtaining the sonar image of seafloor bottom, preprocessing the image with denoising and enhancement, extracting the edge shape based on a canny algorithm, and generating a gray scale-basic element co-occurrence matrix, constructing a convolution neural network classifier structure and a sample set, training a neural network, obtaining classification model and realizing bottom material sonar image classification. The present invention is directed to the graphical characteristics of bottom material sonar images of seafloor, the disadvantage of using single method is solved, and the learning strategyof the convolution neural network classifier is used to learn and train different types of seabed sediment, and finally the classification model with classification function is obtained, and the purpose of fast and accurate classification of seabed sediment sonar images is achieved.

improved LeNet-5 fusion network traffic sign recognition method for assisting driving is disclosed. The method comprises the following steps of constructing a traffic sign image data set and constructing a training set and a test set according to the traffic sign data set, .secondly, constructing a convolutional neural network CNN; substituting the training set image data into a convolutional neural network CNN for training, and storing the trained model; And finally, substituting the test set image data into the trained model for prediction to obtain a prediction result.

The invention discloses an intelligent fault diagnosis method based on a circulation neural network. The method comprises following steps of (1) by use of an acceleration sensor, acquiring time sequence vibration signals of a rotary machine during work in different health states, and dividing acquired original vibration signals into a training set and a test set; (2) establishing a circulation neural network; (3) training the circulation neural network; (4) testing the trained circulation neural network, according to a classification result, judging whether the network reaches an expected diagnosis target, and if the accuracy is lower than the expected value, repeating the step (3) until the circulation neural network whose accuracy is higher than the expected value is acquired; and (5) carrying out intelligent fault diagnosis through the circulation neural network obtained in the step (4). According to the invention, by use of the modeling ability for sequence information of the circulation neural network, the original time sequence vibration signals are directly processed, so faults of the rotary machine can be precisely diagnosed by use of lesser information, and quite high identification speed is achieved.

The invention discloses a data classification model training method and device, and relates to the field of computer technology. First, a universal data set of recognition objects with a small number of classes and a large number of samples in each class is used to allow a classification model to have initial recognition performance. Then, parameters of several layers at the top of the classification model are trained by an actual data set of the recognition objects with a large number of classes, so that the classification model can be adapted to specific recognition scenes and reach a convergence effect. Afterwards, the identification data of the object is used to train the classification model overall, so that the classification model can guarantee the convergence and prevent over-fitting, and thus the performance and accuracy of data classification are ensured.

The invention relates to the technical field of image processing, in particular to a fetus head full-automatic segmentation method based on three-dimensional ultrasound, which comprises the followingsteps of firstly, carrying out data enhancement on a three-dimensional ultrasound volume data set of a fetus head to obtain an enhanced data set; inputting the enhanced data set into a full convolutional neural network, and training a model in an end-to-end volume mapping manner to realize pre-segmentation of the data set; and finally, carrying out iterative optimization processing on a pre-segmentation result by adopting a cascade full convolutional neural network based on an automatic context to obtain a final segmentation result. The invention aims to overcome a plurality of defects in fetal head measurement by the existing two-dimensional ultrasound, so that the subsequent diagnosis efficiency of doctors is improved, and more other prenatal studies are promoted.

The invention discloses a residual network and attention mechanism-based drug relationship extraction method. The method comprises the following steps of: S1, carrying out vector representation on words in a drug entity relationship data set; S2, carrying out time-series modeling on a drug relationship statement by utilizing a two-layer bidirectional long-short term memory model neural network; S3, importing residual connection into the constructed two-layer bidirectional long-short term memory model neural network; S4, decomposing a deep semantic meaning automatically obtained by the two-layer bidirectional long-short term memory model neural network into a memory space and an attention space, fusing memory information and attention information, and inputting the fused information into aSoftmax classifier to extract a drug relationship. According to the drug relationship extraction method disclosed by the invention, dependency relationships between long-distance words are effectivelysolved, gradient diffusion is overcome, model overfitting is prevented, the model robustness is good and the classification effect is good.

The invention relates to a fine-grained image classification method based on a sparse bilinear convolutional neural network, and the method comprises the steps: carrying out the feature channel cutting of the bilinear convolutional neural network, automatically thinning feature channels in a training process, distinguishing the importance of the feature channels for classification, and carrying out the scale cutting according to the importance; inputting the output of the bilinear convolutional neural network into the batch regularization; taking a scaling factor of BN as a scale factor; applying a regularization method to the scale factor, wherein the regularization method has a plurality of types such as L1 and L2, the sparsity of L1 is strong, the sparsity of the feature channels can be realized by jointly training the network weight and the scale factor; finally, performing pruning according to the size sequence of the sparse scale factor, and finally, obtaining a model for finally performing a fine-grained image classification task by utilizing fine tuning. Weak supervision can be realized, redundant parameters are reduced, overfitting is prevented, and the accuracy of fine-grained image classification is effectively improved.

The invention relates to a method, device and system for detecting tire defects. The method comprises the following steps of acquiring an X-ray image of a tire; inputting the X-ray image into a trained Mask R-CNN model to carry out defect identification; outputting an identification graph with the defect marks, wherein the Mask R-CNN model training process comprises the following steps of acquiring the X-ray image of the tire, and performing segmentation of a set pixel size on each image by using an image segmentation software; marking the segmented image through a visual image annotator; andtaking the marked image as a training set to carry out adaptive training on the Mask R-CNN model. Compared with the prior art, the defect detection and the classification can be carried out on the detected image at the same time through the trained Mask R-CNN model, so that the defect detection accuracy is obviously improved.

The invention discloses a tool abrasion state identification method based on a convolutional neural network and long-short-time memory neural network combined model. A force measuring instrument and an acceleration sensor are arranged on a workbench clamp of a numerical control machine tool and a workpiece, three-direction force signals and vibration acceleration signals are collected, collected data are subjected to data pre-processing, normalization processing and unified segmentation are conducted on the same row of data, one-dimension data are converted into two-dimension data to serve asinput, the convolutional neural network in the combined model is used for extracting abstraction features, the long-short-time memory neural network in the combined model is used for finding relevancebetween the data, and finally the tool abrasion state is output. An established double-network structure is arranged in a serial manner, the internal relation between the two kinds of signals can beestablished, the more abstract features are extracted through convolution, the timing sequence feature is determined according to the long-short-time memory, accordingly, the purpose of deeper relation of the data and the model is achieved, and applicability is achieved on various machine tools.

The invention provides a target classification and positioning method based on network supervision. The target classification and positioning method comprises the following steps: automatically obtaining a large amount of network image data from a search engine according to the category of a to-be-tested target; filtering to remove noise images to form a training sample set; preliminarily constructing a classification and positioning network; and inputting samples in the training sample set into the preliminarily constructed classification and positioning network to perform feature extraction,classifying the features, obtaining position information of the target object, and training the classification and positioning network. According to the end-to-end fine classification and positioningmethod based on network supervision, massive network images easy to obtain are used as a training set, manual annotation is completely removed, only image-level labels are used, an efficient convolutional network is designed, and algorithms such as global average pooling and class activation mapping graphs are fused, so that the performance of the method exceeds that of a weak supervised learningmethod on fine classification tasks and positioning tasks.

A method for using artificial neural networks as a universal computing device to model the relationship between the training inputs and corresponding outputs and to solve all problems with estimation, classification, and ranking tasks in their nature. Raw data related to problems is obtained and a subset of that data is processed and distilled for application to this universal computing device. The training data includes inputs and their corresponding results, which values could be continuous, categorical, or binary. The goal of this universal computing device is to solve problems by the universal approximation property of artificial neural networks. In this invention, a practical solution is created to resolve the issues of local minima and generalization, which have been the obstacles to the use of artificial neural networks for decades. This universal computing device uses an efficient and effective search algorithm, Retreat and Turn, to escape local minima and approach the best solutions. Generalization for this universal computing device is achieved by monitoring its non-saturated hidden neurons as related its effective free parameters and In-line Cross Validation process. The output process of ranking is achieved by an added baseline probability retaining from best logistic regression model as a secondary order while the categorical results from a MLP neural network as the first order.

The invention provides a GAN generated picture detection method and a system based on a residual domain rich model, and the method comprises the steps: an original image processing step: carrying outthe recognition and cutting of an original image through employing a digital image processing technology, recognizing a face, and cutting an image of a face part; a step of obtaining residual information of the original image: preprocessing the original image by using a digital image processing technology, and extracting the residual information of the original image; a convolutional neural network processing step: inputting a residual image of the original image into a set convolutional neural network, and adding a BN layer to a convolutional layer in front of each activation function; a global average pooling layer processing step: replacing a full connection layer with a global average pooling layer; and a sample training convolutional neural network processing step: training a convolutional neural network by using the sample of the data set to obtain a picture classifier, and obtaining judgment result information. According to the method, a preprocessing high-pass filter is designed, and 99% of accuracy is finally achieved through an improved convolutional neural network.

Embodiments of the invention provide a method, an apparatus and a device for optimizing a wakeup model, and a storage medium. The method comprises the following steps: acquiring a training set and a verification set; performing iterative training on the wakeup model according to the training set and the verification set; in the iterative training process, periodically updating the training set andthe verification set according to the wakeup model and a preset corpus, and continuing to perform iterative training on the wakeup model according to the updated training set and verification set; and outputting the wakeup model when a preset stopping condition is met. According to the method, the training set and the verification set are updated periodically according to the wakeup model and thepreset corpus in the iteration process, so that the optimization efficiency and effect of the wakeup model can be improved, the stability and adaptability of the wakeup model can be improved, and theoverfitting is avoided.

The invention discloses an epileptic lesion three-dimensional automatic positioning system based on deep learning. The system comprises: a PET image acquisition and marking module; a PET image and standard symmetrical brain template registration module; a PET image data preprocessing module wich is used for generating mirror image pairs of the left and right brain image blocks; a twin network SiameseNet training module which comprises two deep residual convolutional neural networks sharing weight parameters, wherein an output layer is connected with a multi-layer perceptron and a softmax layer, and the networks are trained by using a training set carrying images of epileptic lesions and normal images to obtain a network model; and a classification module and an epileptic focus positioningmodule which are used for generating a probability heat map from a newly input PET image by utilizing the trained network model, judging whether the image is normal or carries an epileptic focus sample through a classifier, and then predicting the position of an epileptic focus area. According to the system, the epileptic focus of the PET image is automatically positioned by introducing the mirrorimage pair of the image block and the twin network SiameseNet, so that the positioning accuracy and efficiency of the epileptic focus can be effectively improved, and the robustness is relatively high.

The invention provides an MRI image inhomogeneity correction method based on regularization. A lowpass filtering method is improved, part of an area with low grey-value difference is selected and is subjected to lowpass filtering to obtain an initial estimate for a inhomogeneous field, and then extrapolation is performed through a regularized polynomial fitting method to obtain an estimate for an inhomogeneous field of a whole image. According to the method provided by the invention, an area-of-interest is determined and is subjected to filtering, so the boundary effect in the global image filtering method can be avoided, and fitting is performed through a regularization method, thus overfitting is guaranteed not to happen under the condition of improving the fitting order.

The invention relates to a construction method and application of a lightweight gesture detection convolutional neural network model, and the method comprises the steps: constructing a lightweight gesture detection convolutional neural network framework based on a SquezeNet convolutional neural network framework and an SSD multi-target detection convolutional neural network framework; Acquiring agesture picture and a background picture, and performing image data enhancement and picture synthesis processing on the gesture picture based on the background picture to obtain a gesture data set; And based on the public data set and the gesture data set, training a lightweight gesture detection convolutional neural network framework to obtain a lightweight gesture detection convolutional neuralnetwork model. According to the invention, a small amount of gesture data is expanded into the gesture data set containing a large amount of picture data at a high speed; The technical problem that alarge amount of high-quality gesture picture data is difficult to obtain is solved, in addition, by combining the SquezeNet convolutional neural network architecture and the SSD multi-target detectionconvolutional neural network architecture, the constructed lightweight gesture detection convolutional neural network model occupies few computing resources, and can be applied to various detection platforms.

A traffic sign recognition method based on deep learning belongs to the field of image recognition. The method comprises the following steps: step 1, acquiring a traffic sign data set, and preprocessing the data set; and step 2, constructing a TSR _ ConvNet network structure for training. And the TSR _ ConvNet network is added into a Dropout strategy for improving overfitting, and is added into abatch normalization layer. And designing a proper convolution kernel size. Wherein the softmax classification layer of the full connection layer adopts a Label-smoothing strategy, and the softmax classification layer of the full connection layer adopts a Label-softing strategy. And 3, inputting the preprocessed traffic sign image into a TSR _ ConvNet network model, and carrying out traffic sign identification. According to the method, the model structure is simplified under the condition of ensuring relatively high accuracy, so that the calculated amount of the model is small, the training time is greatly reduced, and the method has better practicability.

The invention discloses a protein-DNA binding residue prediction method based on sampling and integrated learning. The method comprises the steps of (1) feature extraction and training sample set construction, (2) sampling and model training, (3) model integration, and (4) online prediction. The method is used for solving the shortcomings of low prediction precision caused by the problems of few feature types and class imbalance in protein-DNA binding residue prediction problems and has the advantages of high prediction precision and high generalization ability.

The invention discloses a machine tool cutter residual life prediction method based on LSTM + CNN, and the method comprises the steps: carrying out the judgment of the signal features of uploaded training data, and distinguishing a continuous signal and a discrete signal; performing data merging on the real-time data of different frequencies sampled by the sensor; checking whether missing values or abnormal values exist in the training data and the real-time data or not; if the missing values or the abnormal values exist, using a moving average method to supplement the missing values or replacing the abnormal values, so as to enable the data to be complete and effective, and removing outliers; carrying out selection and dimension reduction on the training data and the real-time data according to data characteristics so as to facilitate model fitting and prevent an over-fitting phenomenon; and training and testing the LSTM + CNN model, and adjusting training parameters and model parameters according to the error, so as to reduce the error to a reasonable range. According to the method, the precision of the prediction result is improved by adopting a grouping mode and a dimension reduction mode, deterministic factors and uncertain factors are comprehensively considered, and the precision of the prediction result can be effectively improved.

The invention discloses a transfer learning-based automatic tomato disease and insect pest detection method, and belongs to the technical field of images. The method comprises the following steps of: collecting tomato disease and insect pest leaf surface picture training samples, carrying out data enhancement on a picture of each tomato disease and insect pest leaf surface picture training sample, and inputting the data enhanced pictures into a deep learning network in a classification manner; and learning the tomato disease and insect pest leaf surface picture training samples, setting a learning frequency upper limit and a recognition precision threshold value, and when a learning frequency achieves the learning frequency upper limit or a recognition precision achieves the recognition precision threshold value, outputting a connection weight model among layers of a current convolutional neural network to serve as a tomato disease and insect pest leaf surface picture recognition classifier.