1443results about How to "Prevent overfitting" patented technology

The present invention is a method of identifying a user's gestures for use in an interactive game application. Videocamera images of the user are obtained, and feature point locations of a user's body are identified in the images. A similarity measure is used to compare the feature point locations in the images with a library of gestures. The gesture in the library corresponding to the largest calculated similarity measure which is greater than a threshold value of the gesture is identified as the user's gesture. The identified gesture may be integrated into the user's movements within a virtual gaming environment, and visual feedback is provided to the user.

The invention discloses an image classification method based on a semi-supervised self-paced learning cross-task deep network. The method includes the steps of randomly selecting a small amount of labeling samples from the whole image data set, reserving the labels, and remaining all the samples as unlabelled samples having the real labels to be unknown in the whole process, wherein the weight ofthe labeled samples is constant to be one in the training process, the weight of the unlabelled samples is initialized to be zero, and only the labeled samples are used as a training set in the initial process; S2, training a cross-task deep network by the training set; S3, according to the trained cross-task deep network, predicting the pseudo labels of all the unlabelled samples, and giving a corresponding weight of each unlabelled sample; S4, according to a self-paced learning normal form, selecting an unlabelled sample with a high confidence degree, and adding to the training set; and S5,repeating the steps S2-S4 until the cross-task deep network performance is saturated or reaches a preset cycle number. According to the method, the human design feature is not needed to be input, andthe classification can be realized by directly inputting the original image.

The invention discloses a human behavior recognition method based on an attention mechanism and a 3D convolutional neural network. According to the human behavior recognition method, a 3D convolutional neural network is constructed; and the input layer of the 3D convolutional neural network includes two channels: an original grayscale image and an attention matrix. A 3D CNN model for recognizing ahuman behavior in a video is constructed; an attention mechanism is introduced; a distance between two frames is calculated to form an attention matrix; the attention matrix and an original human behavior video sequence form double channels inputted into the constructed 3D CNN and convolution operation is carried out to carry out vital feature extraction on a visual focus area. Meanwhile, the 3DCNN structure is optimized; a Dropout layer is randomly added to the network to freeze some connection weights of the network; the ReLU activation function is employed, so that the network sparsity isimproved; problems that computing load leap and gradient disappearing due to the dimension increasing and the layer number increasing are solved; overfitting under a small data set is prevented; and the network recognition accuracy is improved and the time losses are reduced.

The invention discloses a generative adversarial network-based pixel-level portrait cutout method and solves the problem that massive data sets with huge making costs are needed to train and optimizea network in the field of machine cutout. The method comprises the steps of presetting a generative network and a judgment network of an adversarial learning mode, wherein the generative network is adeep neural network with a jump connection; inputting a real image containing a portrait to the generative network for outputting a person and scene segmentation image; inputting first and second image pairs to the judgment network for outputting a judgment probability, and determining loss functions of the generative network and the judgment network; according to minimization of the values of theloss functions of the two networks, adjusting configuration parameters of the two networks to finish training of the generative network; and inputting a test image to the trained generative network for generating the person and scene segmentation image, randomizing the generated image, and finally inputting a probability matrix to a conditional random field for further optimization. According tothe method, a training image quantity is reduced in batches; and the efficiency and the segmentation precision are improved.

The invention relates to an image description generation method based on a depth LSTM network, comprising the following steps: (1) extracting the CNN characteristics of an image in an image description dataset, and acquiring an embedded vector corresponding to the image and describing the words in a reference sentence; (2) building a double-layer LSTM network, and carrying out series modeling based on the double-layer LSTM network and a CNN network to generate a multimodal LSTM model; (3) training the multimodal LSTM model by means of joint training; (4) gradually increasing the number of layers of the LSTM network in the multimodal LSTM model, carrying out training each time one layer is added to the LSTM network, and finally, getting a gradual multi-objective optimization and multilayer probability fused image description model; and (5) fusing the probability scores output by the branches of the multilayer LSTM network in the gradual multi-objective optimization and multilayer probability fused image description model, and outputting the word corresponding to the maximum probability through common decision. Compared with the prior art, the method has such advantages as multiple layers, improved expression ability, effective updating, and high accuracy.

The invention relates to a virtual learning environment micro-expression recognition and interaction method based on a double-flow convolutional neural network, and the method comprises the followingsteps: S1, carrying out the preprocessing of micro-expression data: carrying out the Euler video amplification of a micro-expression video, extracting an image sequence, carrying out the face positioning of the image sequence, and carrying out the cutting of the image sequence, and obtaining the RGB data of a micro-expression; extracting optical flow information from the data amplified by the Euler video to obtain an optical flow image of the micro-expression; s2, dividing the preprocessed data into a training set and a test set, and constructing a double-flow convolutional neural network by using a transfer learning method so as to learn space and time domain information of micro expressions; s3, carrying out maximum value fusion on the output of the double-flow convolutional neural network to enhance the recognition accuracy and obtain a final micro-expression recognition model; and S4, creating a virtual learning environment interaction system by using the micro-expression recognition model, and obtaining a user face image sequence through Kinect to carry out a micro-expression recognition task.

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 polarization SAR image classification method based on CNN and SVM, and mainly aims to solve the problem of the existing polarization SAR image classification method that the classification precision is low. The method comprises the steps as follows: (1) inputting a to-be-classified polarization SAR image after filtering; extracting and normalizing the original feature of each pixel point based on a polarization coherence matrix and by taking the neighborhood into consideration; training an AE network, and obtaining the parameter of a CNN convolution layer through softmax fine-tuning; setting a CNN pooling layer as average pooling, and determining the parameter of the CNN pooling layer; and sending the features of CNN learning to an SVM for classification to obtain the classification result of the polarization SAR image. Compared with the existing methods, the spatial correlation of the image is fully considered, a new neighborhood processing method is proposed based on CNN, features more conductive to polarization SAR image classification can be extracted, the classification accuracy is obviously improved, and the method can be used for polarization SAR image surface feature classification and object identification.

The invention provides a real-time voice conversion method under conditions of minimal amount of training data. The method utilizes an ensemble learning theory to carry out modeling of a Gaussian mixture model to the collected data and design a mapping function under the rule of minimum mean square error. The method solves the problem that a standard GMM easily leads to over-fitting in the case of very minimal amount of data, and increases the robustness of a voice conversion algorithm for amount of data issues. At the same time, the GMM with more standard computational complexity is low in the process of estimating GMM parameters by the method, so the method is suitable for real-time voice conversion.

The invention discloses a motion sensing game interactive method and system based on deep learning and big data, and belongs to the technical field of man-machine interaction. The method includes the steps of I, collecting a motion video sample dataset; II, establishing a deep convolutional neural network model, and training the same offline; III, using the deep convolutional neural network model. The system comprises a deep convolutional neural network offline training module, a real-time man-machine interaction module, and a deep network model online optimization module based on big data. Common cameras are used in real time to acquire game operation videos of game players, high-level semantic characteristics of motion via are extracted via a deep convolutional neural network, limb movements are judged and converted into actual control data for a game target, and the game target can be controlled to move correspondingly the same as the player via the player's limbs.

The invention discloses a polarimetric SAR classification method based on a semi-supervised convolutional neural network. According to the method, first, image data is input, and super pixel segmentation is performed; training samples and test samples are extracted; a neighborhood preserving and sparse filtering deep learning network (NDSFN) is adopted to perform unsupervised pre-training on parameters of the convolutional neural network, and a needed filter set is obtained; a feature map is obtained through convolution operation; fuzzy processing is performed on the feature map through downsampling; parameter settings of the NDSFN are adjusted according to the filter size in a second convolution layer to obtain a new feature map; a Softmax classifier is utilized to perform image classification on the training samples; and a small quantity of marked samples are adopted for slight adjustment to obtain a polarimetric SAR terrain classification. Through the method, by constructing a novelsemi-supervised convolutional neural network (SNCNN) model, the problem that a traditional supervised convolutional neural network has a high requirement for marked samples is effectively solved.

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 discloses a pedestrian re-identification method based on depth learning, includes Step 1: pre-training the CNN model: including pedestrian feature extraction and feature measurement, wherein pedestrian feature extraction adopts the method of combining global feature and local feature, Euclidean distance is used as similarity measure, and loss function based on measure matrix is established under the constraint condition of distance of eigenvector. The loss function is optimized by adding constraint function on the basis of traditional Triplet Loss. 2, testing the data set: inputting the image of the teste data set into the CNN model trained in the step 1, obtaining the image features, calculate the similarity between the target pedestrian image and the reference pedestrian image by the Euclidean distance, and finally arranging the reference pedestrian image according to the similarity magnitude to obtain the pedestrian re-recognition result. The method is suitable for there-recognition of people in complex scenes, and has strong transplantability for scene changes, stable algorithm, fast speed and strong practicability.

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.

A predictive model method (and structure) includes receiving an input data into an initial model to develop an initial model output and receiving both of the input data and the initial model output as input data into a first transform/regression stage.

The invention discloses a wireless positioning method and system based on depth learning and aims to solve a problem of non-high indoor positioning precision. The method comprises steps that S1, wireless signal intensities received by all sampling points are acquired; S2, a depth learning algorithm is utilized to input the acquired wireless signal intensities and coordinates to a depth learning neural network to generate a depth learning neural network with a weight; and S3, wireless signal intensities of users are acquired and are inputted to the depth learning neural network with the weight to determine the positions of the users. The method is advantaged in that indoor positioning precision can be improved, and an over-fitting problem in a full connection layer of the neural network can be avoided.

Bearing fault diagnosis based on pseudo-tag semi-supervised kernel local Fisher discriminant analysis is provided. The method is characterized in comprising the following steps: (1) collecting vibration signals of bearings under different working conditions to form a training sample; (2) performing feature extraction on the training sample obtained in (1); (3) performing normalization processing on features obtained in (2); (4) obtaining a clustering tag set by using density peak clustering for the entire feature set obtained in (3); (5) using clustering pseudo-tags obtained in (4) to construct local inter-cluster divergence and intra-cluster divergence regularization terms, and combining the regularization terms with the inter-class divergence and intra-class divergence with tag samples in the FDA to determine a final projection vector; (6) using the final projection vector obtained in (5) to solve a projection set of the tagged feature set in the dimensionality reduction space; (7) using the projection set obtained in (6) to train an extreme learning machine; and (8) performing processing of (2), (3) and (5) on the collected vibration signals, and inputting the processed vibration signals to determine the working conditions. The technical scheme of the present invention is applied to the problem of fault identification of bearing equipment.

The invention discloses a robust mechanism research method of characteristic significance in image quality evaluation. The robust mechanism research method comprises the following steps: firstly, determining a target function of characteristic selection in the image quality evaluation, and initializing a model parameter; secondly, adding an optimal characteristic into a characteristic matrix, and removing a characteristic disturbance term; thirdly, calculating the significance of the characteristic selection in an image quality evaluation system; fourthly, judging whether the significance meets a system robust requirement or achieves an upper limit of a characteristic number; and finally, verifying a model classification effect. The characteristic significance is measured through an imported system characteristic signal to noise ratio, a constrained optimization problem of a smooth convex function in the image quality evaluation system is solved, interference on a classification face by non-significant characteristics is effectively lowered, the robustness of the image evaluation system is improved, and the self-adaptive optimization problem of characteristic attribute selection on the basis of an image quality evaluation network of a learning mechanism is solved.

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.