59 results about "Max pooling" patented technology

The invention relates to a rapid target detection method based on a convolutional neural network, and relates to the computer vision technology. The rapid target detection method comprises the following steps: training convolutional neural network parameters by utilizing a training set; solving the problem of max-pooling losing feature by using an expander graph and generating a discriminative complete feature graph; regarding the full-connection weight of the convolutional neural network as a linear classifier, and estimating the generalization error of the linear classifier on the discriminative complete feature by using a probable approximately correct learning framework; estimating the required number of the linear classifiers according to the generalization error and the expected generalization error threshold value; and finally, completing the target detection on the discriminative complete feature graph by using the linear classifiers on the basis of a smooth window. The detection efficiency and the target detection precision are obviously improved.

An in vivo face detection method based on a convolutional neural network relates to the machine learning and mode identification field, and is provided to face a cheat problem in the face identification. A conventional face identification technology is very easy to attack, the attackers frequently copy the faces of the legal users by the methods, such as the photographs, the videos, the 3D models, etc., if a face identification system can not distinguish the real faces and the fake faces effectively, the invaders can pass the face identification system very easily by the fake identities. Based on the problem, the present invention provides the in vivo face detection method based on the convolutional neural network. The convolutional neural network in the method provided by the present invention is realized based on a cuda_convnet framework, the network structure comprises four convolutional layers of two max-pooling layers, one full connection layer and one soft_max layer, and the soft_max layer comprises two nerve cells used for predicting the probability distribution of the real and fake faces.

The invention provides a lightweight and efficient neural network model, and designs a multi-task scene semantic comprehension model based on the model. The neural network comprises the following fivestages that: stage one: adopting an mC.ReLU module; stage two: adopting Max-Pooling to carry out downsampling, and then, adopting a plurality of Block-B modules; stage three: adopting a plurality ofBlock-B modules and a plurality of Block-D modules; stage four: adopting a plurality of Block-A modules and a plurality of Block-C modules; and stage five: adopting a plurality of Block-A modules. Themulti-task scene semantic comprehension model comprises the neural network model, a multilevel fusion network and a multi-task semantic prediction network, wherein the neural network model and the multilevel fusion network share different tasks to serve as a common feature extractor. The model can be applied to the field of automatic driving scene perception, and is favorable for realizing road obstacle detection, scene semantic segmentation and real-time multi-task semantic prediction.

A method re-identifies objects in a pair of images by applying a convolutional neural network (CNN). Each layer in the network operates on an output of a previous layer. The layers include a first convolutional layer and a first max pooling layer to determine a feature map, a cross-input neighborhood differences layer to produce neighborhood difference maps, a patch summary layer to produce patch summary feature maps, a first fully connected layer to produce a feature vector representing higher order relationships in the patch summary feature maps, a second fully connected layer to produce two scores representing positive pair and negative pair classes, and a softmax layer to produce positive pair and negative pair probabilities. Then, the positive pair probability is output to signal whether the two images represent the same object or not.

Feature extraction, coding and pooling, are important components on many contemporary object recognition paradigms. This method explores pooling techniques that encode the second-order statistics of local descriptors inside a region. To achieve this effect, it introduces multiplicative second-order analogues of average and max pooling that together with appropriate non-linearities that lead to exceptional performance on free-form region recognition, without any type of feature coding. Instead of coding, it was found that enriching local descriptors with additional image information leads to large performance gains, especially in conjunction with the proposed pooling methodology. Thus, second-order pooling over free-form regions produces results superior to those of the winning systems in the Pascal VOC 2011 semantic segmentation challenge, with models that are 20,000 times faster.

An improved method of performing object segmentation and classification that reduces the memory required to perform these tasks, while increasing predictive accuracy. The improved method utilizes a capsule network with dynamic routing. Capsule networks allow for the preservation of information about the input by replacing max-pooling layers with convolutional strides and dynamic routing, allowing for the reconstruction of an input image from output capsule vectors. The present invention expands the use of capsule networks to the task of object segmentation and medical image-based cancer diagnosis for the first time in the literature; extends the idea of convolutional capsules with locally-connected routing and propose the concept of deconvolutional capsules; extends the masked reconstruction to reconstruct the positive input class; and proposes a capsule-based pooling operation for diagnosis. The convolutional-deconvolutional capsule network shows strong results for the tasks of object segmentation and classification with substantial decrease in parameter space.

The invention provides an image recognition method based on biological vision and a precise pulse drive neural network. The method is enlightened by a biological vision layering system, on a feature extraction part of an image, an HMAX model is adopted for simulating and sensing a wild cell running mechanism, Gabor filtering is firstly utilized for enhancing marginal information of an image, images obtained after Gabor filtering in various directions are subjected to max pooling processing, and the aims of extracting major features and carrying out dimension reduction processing are achieved. According to the feature image data processing method, a phase encoding method is selected, pixel information of the images is converted into a pulse phenomenon, thus, space information of the images is considered, and time information of the images is also considered. The method has certain biological basis, and has good feasibility and robustness, and the accuracy of image recognition and classification, especially in noise images is greatly improved.

The invention relates to a multi-modal image target detection method based on image fusion, and the method comprises the steps: 1), making a multi-modal image data set through an infrared image and avisible light image which are collected in advance; 2) taking the preprocessed paired images as the input of a generation model G in the fusion model; generating a model G based on U-Net full convolutional network. A convolutional neural network based on a residual network is used as a generative network model structure and comprises a contraction process and an expansion process, a contraction path comprises a plurality of convolution plus ReLU active layer plus maximum pooling (Max Pooling) structures, the number of feature channels of each step of down-sampling is doubled, and a generated fusion image is output; the fused image is input into a discrimination network model in a fusion model; according to the change of a loss function in a training process, a learning rate training indexis adjusted according to the number of iterations, and through training, based on a self-owned multi-modal image data set, an image fusion model which retains infrared image thermal radiation characteristics and visible light image structural texture characteristics at the same time can be obtained.

The invention relates to the in-depth learning field, and especially relates to a Convolution Neural Network (CNN) structure improving method; the method comprises the following steps: a, using a fractional max pooling (FMP) principle to change a maximum value pooling layer in a conventional CNN structure into fractional orders, thus realizing sampling dimension reduction under a random dimension; b, ensuring a shallow network structure, continuously widening the network structure, combining the fractional order maximum value pooling layer, and thus improving the network performance. The method uses the fractional order maximum value pooling principle to ensure the layer at a shallow level, can continuously widens the network structure, thus preventing a deep network to have gradient vanish and weight failure phenomenon in a training process, and causing the CNN structure hard to train. The method can realize equal or better performance with the deep CNN structure, and uses less network parameters, thus providing obvious performance advantages.

The present invention provides a dense three-dimensional object reconstruction method based on learning efficient point cloud generation. The main content of the method comprises: an object reconstruction system, structure parameter setting and a training method. The process of the method comprises: an encoder is employed to extract an input image to an abstract representation space and perform union two-dimensional projection, a structure generator is employed to generate point cloud on the basis of two-dimensional convolution; up-sampling operation is employed to enlarge the image resolution, after the point cloud is projected to a two-dimensional plane, and the max-pooling operation is employed to recover the image to the original resolution to obtain a reconstruction object view. The dense three-dimensional object reconstruction method based on learning efficient point cloud generation can process three-dimensional views with different visual angles, and a pseudo compensation device is improved to solve the pixel superposition problem so as to improve the recognition rate of object reconstruction and the generation correct rate in a new perspective.

The invention provides an electronic medical record entity relationship extraction method based on a shortest dependency subtree. The method comprises the following steps: firstly, extracting an entity-based shortest subtree from an original sentence through dependency syntactic analysis to compress the sentence length; secondly, coding the statements through a bidirectional long short-term memory(BLSTM) neural network, and then coding the statements through the BLSTM neural network; learning final semantic representation of the sentences through a maximum pooling layer (Max Pooling), and finally classifying the sentences through a softmax classifier to obtain an entity relationship. According to the method, noise vocabularies and compressed statement lengths can be deleted. Meanwhile, the key words representing the relations between the entities are completely reserved, so that the compressed statement semantic relations are clearer. The problem that semantic information of statements cannot be well represented due to too long statements of an existing electronic medical record entity relation extraction model is solved, and the performance of the relation extraction model is improved.

The invention discloses a method, for labeling object categories which pixels belong to, based on super pixel level features and belongs to the field of scene understanding and image semantic segmentation. Labeled training set images are used for training, and a to-be-labeled image is labeled by using the steps of firstly, extracting and encoding the features of each pixel; secondly, acquiring super pixels through 'over-segmentation; building different-level neighborhood areas for each super pixel according to different distances; for each area, using Max-Pooling comprehensive features to connect the features of different levels so as to obtain the level features of the super pixels; thirdly, using the level features to build and solve random field models based on the super pixels through linear SVM to obtain image pixel labels. By the method, pixel labelling time can be reduced, and labeling precision can be increased. In addition, the CUDA parallel calculation technology can be used during labeling so as to further increase labeling speed.

The invention provides a text sentiment classification method and a device, equipment and a storage medium. The method comprises the steps of obtaining a word vector in to-be-processed text data, andextracting a feature vector corresponding to the word vector; extracting context feature representation of the feature vector by adopting a bi-directional long-short term memory network Bi-LSTM model;according to the extracted context feature representation, utilizing an Attention mechanism, then introducing a top-k-max pooling processing mode to fully extract text feature representation, and theextracted features are sent to a classifier to obtain higher accuracy. According to the method, the classification accuracy of text sentiment classification is improved, and the classification effectis good.

A computer-implemented method, computer program product, and computer processing system are provided for word embedding. The method includes receiving, by a processor device, a word embedding matrix. The method further includes generating, by a processor device, an average pooling vector and a max pooling vector, based on the word embedding matrix. The method also includes generating, by the processor device, a prediction by applying a Multi-Layer Perceptron (MLP) to the average pooling vector and the max pooling vector.

The invention discloses a short text similarity calculation method based on multi-dimensional convolution characteristics, which comprises the following steps: constructing a multi-granularity convolution neural network model by using training data; Two training samples are inputted into the input layer of the multi-granularity convolution neural network model to obtain their word vector matrices.Multi-granularity convolution operation is carried out in the convolution layer to extract respective feature vectors. Using the K-Block-Max pooling and average pooling method in a pooling layer to extract quadratic eigenvectors. In the similarity calculation layer, the similarity vectors of the two training samples are obtained by using the fusion direction and distance calculation method. The similarity values of the two training samples are calculated in the whole connection layer and compared with the similarity values labeled in the training data to update the model. Two pieces of shorttext which need to be calculated similarity are input into the trained multi-granularity convolution neural network model, and the similarity value is output at all connection layers. The invention adopts different granularity convolution check short text data for feature extraction to improve accuracy.

The invention discloses an image classification method based on a field programmable gate array (FPGA). The method comprises the realization steps of (1), obtaining a test data set; (2), obtaining a training data set; (3), establishing a 15-layer convolutional neural network comprising 10 convolutional layers, 3 max-pooling layers, an average-pooling layer and a soft max layer; (4), setting kernelparameters of each layer in the convolutional neural network; (5), training the convolutional neural network through utilization of the training data set; (6), inputting test images into the convolutional neural network, and classifying the test data set; and (7), computing accuracy of the test data set. The method can be used for classifying optical images, SAR (Synthetic Aperture Radar) imagesand natural images on the field programmable gate array (FPGA).

An artificial neural network architecture is provided for processing raw audio waveforms to create speaker representations that are used for text-independent speaker verification and recognition. The artificial neural network architecture includes a strided convolution layer, first and second sequentially connected residual blocks, a transformer layer, and a final fully connected (FC) layer. The strided convolution layer is configured to receive raw audio waveforms from a speaker. The first and the second residual blocks both include multiple convolutional and max pooling layers. The transformer layer is configured to aggregate frame level embeddings to an utterance level embedding. The output of the FC layer creates a speaker representation for the speaker whose raw audio waveforms were inputted into the strided convolution layer.

The invention discloses a dynamic soft measuring method of 4-CBA content based on a convolutional neural network. The dynamic soft measuring method of the 4-CBA content based on the convolutional neural network is used for calculating the content of 4-CBA generated in the PTA oxidation process. The dynamic soft measuring method of the 4-CBA content based on the convolutional neural network comprises the following steps: firstly, constructing a mapping relation between an input and an output of a dynamic soft measurement model on the basis of the convolutional neural network, using a time sequence data block of a relevant measurable variable in the PTA oxidation process as the input of the dynamic soft measurement model and using the 4-CBA as the output of the dynamic soft measurement model; secondly, inputting the time sequence data block into the convolutional neural network in which convolutional layers and pooling layers are alternately distributed, wherein the layer numbers of theconvolutional layers and the pooling layers are both 2, the first layer of pooling adopts characteristics after convolution is extracted in a one-dimensional max-pooling manner, and the second layer of pooling adopts max-pooling equivalent to the sizes of characteristic graphs output by the convolutional layers to perform sampling; calculating the output of the last layer of pooling by using a linear function to obtain an output result; and comparing the result with 4-CBA analysis data and updating parameters. The dynamic soft measuring method of the 4-CBA content based on the convolutional neural network, disclosed by the invention, has the advantages that the dynamic soft measurement model is simple and easy to realization, and the measurement accuracy of the model is improved.

The invention discloses a malicious software operation code analysis method based on a convolutional neural network. The method comprises the steps of: obtaining a Dalvik byte code; obtaining an operation code sequence, and representing the operation code sequence by a one-hot vector; converting the one-hot vector into a vector with a fixed size, multiplying the vector by a random weight matrix, and inputting the vector into a convolutional neural network; outputting a feature mapping set matrix C in the convolution layer; in k-max pooling, performing maximum merging operation on the matrix C,and extracting the most important k characteristic values to output a characteristic vector Z; forming a full connection layer by the vector Z, and operating the vector Z in the full connection layerto obtain an output feature y; processing the output feature y by using a softmax function to obtain relative probability distribution p; calculating a cross entropy loss function Lk; gradually adjusting the minimum loss function and the parameter values of the corresponding model by using a gradient descent method; iteratively updating model parameters based on the output calculations and optimizing the detection model. The method has the characteristic of high detection accuracy.

A method is provided for image processing. The method includes receiving an image dataset having image data for a plurality of images and a group dataset having group information for each image. The method further includes transforming the image dataset into a sparse vector by applying sparse coding and using the group information. The method also includes dividing each of dimensions of the sparse vector into multiple groups with a matrix. The matrix is generated such that the matrix takes on a higher value when samples in a same group from among the multiple groups have same non-zero dimensions above a threshold amount of time and takes on a tower value compared to the higher value when the samples in the same group lack the same non-zero coefficients above the threshold amount of time. The method additionally includes transforming the sparse vector into a summation vector using the multiple groups.

The invention discloses a SAR image classification method based on SAR-SIFT and DBN. The problem of synthetic aperture radar image classification is mainly solved. The classification process comprises the following steps: (1) inputting an integer type SAR image matrix; (2) transforming the SAR image matrix; (3) extracting SAR-SIFT features; (4) max pooling; (5) normalization; (6) training four restricted Boltzmann machines RBM; (7) training a softmax classifier; (8) building a deep belief network DBN; (9) classification; and (10) calculating the classification accuracy. According to the invention, the SAR-SIFT features of synthetic aperture radar images are extracted, the deep belief network DBN is adopted, the features are learned layer by layer, the information integrity of the radar images is retained, depth information is mined, and a good classification effect is achieved. The method can be used for SAR image classification.

The invention provides a deep learning-based lightweight algorithm for recognizing electrocardio data. The algorithm comprises the following steps of: roughly extracting data features of extracted electrocardio data through a standard convolution layer; sending the roughly extracted data features into a core module Lite Module so as to extract deep data features after the roughly extracted data features pass through a pooling layer Max-Pooling; sequentially sending the deep data features into a two-layer full-connection layer dense to purify the deep data features after the deep data featurespass through a pooling layer Max-Pooling; and sending the purified data features into a classifier function to carry out feature classification and output. Compared with other algorithms, the algorithm does not need many calculation parameters while ensuring a certain recognition effect, and is capable of processing electrocardio data on limited network resources or operation memories.

In an optical character recognition (OCR) method for digitizing printed text images using a long-short term memory (LSTM) network, text images are pre-processed using a stroke-aware max-min pooling method before being fed into the network, for both network training and OCR prediction. During training, an average stroke thickness is computed from the training dataset. Stroke-aware max-min pooling is applied to each text line image, where minimum pooling is applied if the stroke thickness of the line is greater than the average stroke thickness, while max pooling is applied if the stroke thickness is less than or equal to the average stroke thickness. The pooled images are used for network training. During prediction, stroke-aware max-min pooling is applied to each input text line image, and the pooled image is fed to the trained LSTM network to perform character recognition.

The invention provides a system and method for completing a three-dimensional laser radar point cloud target. The global features of the laser radar target are obtained by three layers of shared multilayer perceptron and three point-by-point max-pooling layers, so that the object point cloud obtained by the laser radar through scanning is denser and more uniform in distribution, and the details ofthe contour of an object are more complete, so as to realize the purposes of detection recognition and measurement calculation.

The invention discloses a SAR image target recognition method based on a denoising self-coding network. The method comprises the following steps: denoising an image by using a three-dimensional modulematching algorithm (BM3D); for the original image and the denoised training image, the dense feature points are extracted by the sliding window and stored in pairs according to the corresponding positions, which are used for the input of the depth denoising encoder in the later training. Using the SIFT feature extracted from the denoised image as the original input x. The SIFT extracted from thecorresponding undenoised image is used as the noisy input to train the deep denoising self-coding network. The spatial pyramid model (SPM) is used to compute the image feature vector expression, and the maximum pooling (Max Pooling) is used to summarize the local features to obtain the final image description vector. A large number of dense SIFT features are used for deep network training, and thehigh-level representation of deep denoising and self-coding features is used for learning. Finally, the linear support vector machine is trained to classify and recognize the features instead of theinitial local features.

The present invention is suitable for the image classification, and provides a classification method of the hyperspectral images. The classification method comprises the steps of dividing the hyperspectral images into a training set and a test set, extracting the local feature points, using a K-means algorithm to calculate the local feature points of the training set to form a dictionary, adoptinga KNN algorithm to form the nearest neighbor words for the to-be-classified local feature points of the test set in the dictionary, searching the nearest neighbor feature points for the to-be-classified feature points of the test set images, searching the neighbor word having the shortest spectral dimension distance in the nearest neighbor words, introducing the triple constraints of the neighborfeature point, the neighbor words and the spectral dimension distance, solving a constraint least absolute fitting problem to obtain a coding coefficient, pooling the coding coefficient by a max-pooling algorithm, and classifying the test set by taking the obtained coding coefficient as a feature descriptor of the hyperspectral images. According to the present invention, an indeterminacy problemwhen a mapping relationship of the hyperspectral image feature points and the dictionary words is established is solved,and the discernment of the similar images is improved.

The invention relates to a Dropout regularization method based on the sensitivity of activation values. The method is used for image classification. A data training stage of the method comprises the following steps: 1) data preparation: collecting different types of images and marking the image types as labels; 2) structure design: setting a deep convolutional neural network structure; 3) initialization: (1) determining the weight of a convolution filter, initializing the parameters by using a random initialization method and setting the number of times of iteration and (2) setting a probability density function selected in Dropout; 4) forward computing: performing computing layer by layer from the first layer to the last layer, determining the probability of zero setting of each feature point via the probability density function after the max pooling layer, generating a random number between 0 and 1 by using a uniform distribution function, comparing the random number with the probability of zero setting of each feature point, zero-setting the activation value of the feature point if the random number is less than the probability and maintaining the activation value of the featurepoint if the random number is equal to or greater than the probability; 5) back propagation.

The invention discloses a decision tree classification foreign Chinese difficulty evaluation method based on svm and bert model random subspace feature selection. The method comprises the steps: generating 86 statistical features according to the length and readability of an article, carrying out the classification through svm, and obtaining a confidence degree 1; and classifying the encoding features by using svm to obtain a confidence coefficient 2. And fusing the obtained two confidence coefficients to serve as new features, and classifying the features by using a decision tree. And for theencoding feature data, outputting an information result through the-1 layer of the encoding extracted by the BERT model, and then performing average-> max pooling processing to obtain 768-dimensionalfeatures in total without normalization. According to the method, the problems of low efficiency and under-fitting of a traditional algorithm are avoided, and all information is used most reasonably,so that the classification basis is increased, and the effect is remarkable. According to the method, the accuracy of evaluating the difficulty of foreign Chinese is 85.6%.

A feature extractor for a convolutional neural network (CNN) is disclosed, wherein the feature extractor is deployed on a member of the group consisting of (1) a reconfigurable logic device, (2) a graphics processing unit (GPU), and (3) a chip multi-processor (CMP). A processing pipeline can be implemented on the member, where the processing pipeline implements a plurality convolution layers for the CNN, wherein each of a plurality of the convolutional layers comprises (1) a convolution stage that convolves first data with second data if activated and (2) a sub-sampling stage that performs a member of the group consisting of (i) a max pooling operation, (ii) an averaging operation, and (iii) a sampling operation on data received thereby if activated. The processing pipeline can be controllable with respect to which of the convolution stages are activated / deactivated and which of the sub-sampling stages are activated / deactivated when processing streaming data through the processing pipeline. The deactivated convolution and sub-sampling stages can remain instantiated within the processing pipeline but act as pass-throughs when deactivated. The processing pipeline performs feature vector extraction on the streaming data using the activated convolution stages and the activated sub-sampling stages.

The invention discloses a terrestrial-air conversation recitation semantic automatic verification method based on BiLSTM. The method comprises the following steps: making an air-ground communication recitation corpus; Preprocessing the corpus; The word vector sequences of the two statements in each recitation statement pair are input into two parallel BiLSTMs respectively, and outputs of the two statements at all moments are obtained; Matching the output vectors of the two BiLSTMs at each moment to generate a semantic matching matrix between the recitation statement pairs; K-Matching Matrix For Semantic Matching Max pooling processing is carried out to obtain first k matching features, and the matching features form a significant matching feature vector; And judging whether the two statements in the recitation statement pair are consistent or not. According to the method, the semantic feature advantage that the BiLSTM network can automatically learn the statement sequence is utilized,the expenditure of manually designing features can be avoided, and the BiLSTM network can extract more abundant context semantic features.