92 results about "Similarity learning" patented technology

The invention relates to a specific modal semantic space modeling-based cross-modal similarity learning method. The method comprises the following steps of: 1, establishing a cross-modal database which comprises data of multiple modal types, and dividing data in the database into a training set, a test set and a validation set; 2, aiming at each modal type in the cross-modal database, constructinga semantic space aiming at the specific modal, and projecting data of the other modal types to the semantic space so as to obtain a cross-modal similarity aiming at the specific modal; 3, fusing cross-modal similarities, aiming at specific modals, obtained from different modal semantic spaces to obtain a final cross-modal similarity; and 4, calculating a similarity between a query example and a query target by taking any modal type in the test set as a query modal and taking another modal type as a target modal, and obtaining a related result list of data of the target modal according to thesimilarity. According to the method, the correctness of cross-modal retrieval can be improved.

The invention discloses an image retrieval method based on variable-length depth hash learning and mainly relates to the field of image retrieval and depth learning. According to the method, learning of hash codes is modeled into the process of similarity learning. Specifically, the method utilizes a training image for generating a batch of ternary image sets. Each ternary image set comprises two images with the same label and one image with the different label. The purpose of model training is to space image pairs matched to the maximum and the unmatched image pairs in hamming space. The depth convolution nerve network is introduced into the learning part of the method, the image characteristics and hash functions are optimized in a combined mode, and the end-to-end training process is achieved; hash codes output by the convolution network have different weights. For the different retrieval task, a user can regular the length of the hash codes by disconnecting unimportant bits; meanwhile, according to the method, the discrimination of the hash codes can be kept effectively under the circumstance that the hash codes are short.

The invention discloses a multi-class image semi-supervised classifying method and system. The method comprises the steps that firstly, similarity learning is conducted on image samples with tags and image samples without tags in a training set, and similar neighbor images and normalized weights are constructed and used for representing sample similarities; secondly, a class tag matrix is initialized, L2,1-norm regularization is introduced to effectively reduce the influence of mixed signals in prediction tags F of flexible class tags on results, constrains which are not negative and are one in column sum are applied to F at the same time, and thus it is ensured that estimated flexible tags meet the probability definition and non-negativity; finally, parameters are used for balancing the influences of similarity measurement, initial class tags and L2,1-norm regularization on classification, semi-supervised learning modeling is completed, the maximum value of similarity probabilities is taken to be used for image class identification, and classification results are obtained. Due to the fact that the L2,1-norm regularization is introduced, the influence of the mixed signals on the classification is reduced, and thus the classification accuracy is improved. In addition, data outside the training set can be effectively classified, and the expansibility is good.

Methods and systems for improved license plate signature matching by similarity learning on synthetic images comprise generating a plurality of synthetic license plate images; applying one or more transformations to the synthetic license plate images to cause the synthetic license plate images to more closely resemble authentic license plate image captures; and providing the synthetic license plate images as inputs to a machine distance learning algorithm in which weighted similarity scores are calculated between signatures of analogous and non-analogous license plate images and one or more sets of signature weights are iteratively adjusted to increase the likelihood that comparing analogous license plate images results in high weighted signature similarity scores and comparing non-analogous license plate images results in low weighted signature similarity scores.

The invention discloses a feature bag image retrieval method based on a Hash binary code. The method comprises steps that, a vision term list is established; tf-idf (term frequency-inverse document frequency index) weight quantification of vision terms is carried out; vision term characteristic quantification of an image is carried out; an inverted index is established; a feature binary code projection direction is learned; feature binary code quantification is carried out; candidate image sets are retrieved. According to the method, the index is established for an image database, rapid image retrieval is realized, and retrieval efficiency is improved, moreover, through a binary code learning method having the similarity retention capability, the binary code is learned from spatial distance similarity and meaning distance similarity as signature, and image retrieval accuracy is improved. The feature bag image retrieval technology based on the Hash binary code has properties of high efficiency and accuracy, and relatively high use values are realized.

Techniques are provided for incorporating human or user interaction in accordance with the design and / or performance of data mining applications such as similarity determination and classification. Such user-centered techniques permit the mining of interesting characteristics of data in a data or feature space. For example, such interesting characteristics that may be determined in accordance with the user-centered mining techniques of the invention may include a determination of similarity among different data objects, as well the determination of individual class labels. These techniques allow effective data mining applications to be performed in accordance with high dimensional data.

The invention provides a convolutional neural network filter pruning technology based on similarity learning. The method comprises the steps of obtaining the similarity between different filters through similarity survey; performing clustering on the filters according to the similarity; enabling the filters in the same cluster to become more and more similar through convergent training; and finally pruning the filters after the convergent training. The technology solves the problem that the model representation capability is reduced and needs to be re-trained in a convolutional neural networkfilter pruning process, so that the representation capability and precision of a network can be better retained; and in addition, the precision loss of the convolutional neural network is hardly caused, the precision of a model does not need to be restored through re-training, and the better balance of model precision and efficiency is achieved.

The present invention provides a target tracking method based on a long short-term memory network, and relates to the computer vision technology. The method comprises the steps of: performing pre-estimation of a candidate target state by employing a fast matching method based on similarity learning, screening out high-quality candidate target states, and performing classification of the high-quality candidate target states by employing a long short-term memory network. The long short-term memory network comprises a convolutional layer used for feature extraction and a long short-term memory layer used for classification. The convolutional layer is obtained through offline training on a large-scale image data set ILSVRC15 to avoid a risk of overfitting of the target tracking data set. The long short-term memory layer is obtained through online learning and fully employs the time correlation included by an input video sequence so as to have good capacities on adaption of target forms andmotion change. The speed is observably improved, the long short-term memory network capable of adapting target change is utilized to target tracking.

The invention discloses a deep learning based video retrieval method which mainly comprises the following parts: performing video pre-processing by using a convolutional neural network; extracting feature vectors of the video after pre-processing by using a long short-term memory; and finally learning by a similarity learning algorithm to obtain a distance calculation method, and performing similarity calculation according to the method and ranking, so as to obtain a video retrieval result. According to the method disclosed by the invention, scene segmentation and key frame selection are performed by the convolutional neural network, and high-level semantics representing video is extracted, so as to acquire a proper quantity of key frame sequences, and effectively avoid false detection andmissing inspection in shot segmentation. According to method disclosed by the invention, time-order characteristics of the video are extracted by the long short-term memory, so as to obtain a more accurate retrieval result. By similarity learning and a text based matching method, matching accuracy of a similarity measurement method can be promoted. By adopting the method disclosed by the invention, accurate retrieval on a large scale of videos can be realized.

The invention discloses a human face recognition method and device based on tensor description. Firstly, similarity learning is conducted on image samples with labels and samples to be classified and without labels, and similar adjacent figures and normalized weights are configured to represent sample similarity; secondly, a category label matrix is initialized manually; thirdly, to effectively achieve direct induction of human face images outside samples, a regularization term capable of conducting direct induction on the images outside the samples and based on the tensor description is integrated into an existing label propagation model; finally, system modeling is finished through influences of parameter weighing similarity measuring, initial category labels and the regularization term based on the matrix pattern on the human face recognition; the maximum value of the probability of similarity in system output is taken to be used for conducting category identification of the human face images, and the most accurate system recognition result is obtained. By introducing the concept of tensor description, the topological structures among image pixels can be effectively maintained in the induction process of the human face images outside the samples, and the system expansibility is good.

The invention discloses a nuclear norm regularization based low-rank image characteristic extraction identification method and system. Firstly an original training image is subjected to similarity learning to construct a reconstruction weight coefficient; and secondly a nuclear norm measurement based neighborhood reconstruction error is minimized and a projection matrix is subjected to nuclear norm regularization processing to obtain a low-rank projection matrix capable of directly extracting two-dimensional image characteristics, so that the topological structure and correlativity among image pixels can be effectively kept. In addition, it can be ensured that low-rank salient image characteristics are obtained by optimization. An original test image is directly embedded in the low-rank projection matrix obtained by training, low-rank salient characteristics of the image are output, classification is performed by utilizing a nearest neighbor classifier based on low-rank salient characteristics in a training set, and category labels of training image samples with highest characteristic similarity with test image samples are obtained, thereby finishing the classification of the test image samples. By introducing nuclear norm regularization, the robustness of noises in a characteristic extraction process can be effectively ensured and the system performance is better.

The invention relates to a transformer fault diagnosis method based on improved semi-supervised classification of a graph. The method comprises the following steps that content data of characteristic gases of a transformer in various running states is acquired as a fault diagnosis sample set and normalized; a labeled sample and an unlabeled sample are selected from each class of samples for similarity learning, a similarity neighbor graph is constructed, and a weight matrix is calculated; an initial label matrix is constructed; the initial label matrix is standardized; semi-supervised classification modeling is conducted according to a fuzzy neighbor label propagation algorithm to generate a training model; class labels are allocated to the label-free test samples through the training model, and a diagnosis result is obtained. The method is applied to fault diagnosis of the transformer with the unbalanced sample set, and the high fault diagnosis precision can be achieved.

The invention discloses a new cell type identification method based on similarity learning and enhancement thereof. The method designs a new global similarity calculation method, combines with other three conventional local similarity information, screens the genes, and carries out enhancement processing on the global similarity with sparse properties. According to the method, a global similaritycalculation method different from the traditional calculation of the local point-to-point similarity is used, the gene selection and the similarity enhancement are carried out by combining multiple different similarities including the global similarity and the local similarity, and a similarity matrix rich in information is obtained. According to the method, the influence of the factors, such as the technical noise, the biological noise, etc., carried by the single cell data can be effectively reduced, and the type of the single cell can be more accurately identified.

An initial noisy graph topology is obtained and an initial adjacency matrix is generated by a similarity learning component using similarity learning and a similarity metric function. An updated adjacency matrix with node embeddings is produced from the initial adjacency matrix using a graph neural network (GNN). The node embeddings are fed back to revise the similarity learning component. The generating, producing, and feeding back operations are repeated for a plurality of iterations.

The invention provides a thesis recommendation method based on a multilayer drawing, which comprises the following steps: Step I, taking a target thesis and candidate theses as nodes in the drawing, constructing a thesis drawing according to reference relationships among the target thesis and the candidate theses and among the candidate theses; Step II, calculating weights of words in the theses by a word weight calculation method, removing the words with the weights less than a threshold, taking the rest words as the nodes, constructing a keyword feature map by virtue of semantic dictionary information; Step III, formulating a target equation by a similarity learning method based on the multilayer drawing; and Step V, optimizing a target function by a heuristic search algorithm of the map, updating weights of sides in the multilayer drawing, and recommending the thesis according to a final multilayer drawing result. The method allows a computer to be capable of recommending the target thesis according to the interested theses read or marked by a user.

The invention provides an image matting method and system. The image matting method comprises the steps of 1, training an image matting model by using an image training set; wherein samples in the image training set comprise an original image, and a tripartite graph, a gold standard mask and a gold standard combined mask which correspond to the original image; wherein the image matting model comprises a depth feature extraction module which is used for learning semantic features and detail information features of an original image; The similarity learning module is used for fusing the semanticfeatures and the detail information features to obtain a similarity relation of the pixel points; The mask propagation module is used for obtaining a mask value of each pixel point through a propagation algorithm according to the similarity relation between the trisection image and the pixel points, and outputting an alpha mask image corresponding to the original image; and 2, inputting the original image to be subjected to image matting and the corresponding tripartite image into the trained image matting model, and outputting an alpha mask image corresponding to the original image to be subjected to image matting.

The present invention provides an image semantic segmentation method adopting inverse attention learning and pixel similarity learning. The method comprises the following steps that: step A, preliminary semantic segmentation is performed on input images, and the branch networks of different scales in a DeepLab v2 ResneT1 101 network are adopted to extract the features of the input images of different scales; step B, on the basis of the step A, an inverse attention layer is adopted to segment the boundaries of the input images; step C, on the basis of the step A, a pixel similarity learning layer is adopted to further segment the boundaries of the input images; step D, the inverse attention layer and the pixel similarity learning layer are optimized, and a corresponding loss function is defined; and step E, network parameters are trained. According to the image semantic segmentation method of the present invention, the inverse attention mechanism is adopted to correct boundary locatingbetween a target area and a background area; the pixel similarity learning mechanism is adopted to solve the problems of boundary locating ambiguity and boundary smoothing between the target area andthe background area; and therefore, the effective segmentation of a fused area between the target area and the background area can be realized.

An object identification apparatus selects, for each partial area of an object in a registered image, a corresponding sample image from sample images based on feature quantities of the partial area, for objects in the registered image, sets a similarity of feature quantities for each partial area between objects of an identical individual and between objects of different individuals based on a similarity related to the selected sample images for each partial area, makes a discriminator learn based on the set similarity, acquires a similarity for each partial area between objects in an input image and the registered image, and determines whether the object in the input image is identical to the object in the registered image based on the acquired similarity and a result of discrimination by the discriminator.

The invention discloses a human face set matching method and system based on similarity learning. Training set samples and testing set samples are selected from original data samples, training samples are selected, the true similarity is calculated and is compared with the worked out calculation similarity, and then a final classifier is selected; the geometric mean of each type of training samples and all of testing samples in the testing set samples are substituted into the final classifier, classifying results are obtained, and then the categories of the testing samples are obtained. According to the scheme, part of samples are selected as the training samples to carry out training, classifiers are selected, and therefore it is avoided that all samples serve as the training samples to be trained, the training process is simplified, the complex process is avoided, and the training speed is improved. In addition, according to the scheme, the geometric mean of each type of samples in the training set samples is selected to construct a plurality of different classifiers, and the effect of the accurate result through the simple operating process is achieved.

The invention provides an image segmentation method based on correlation matrix self-learning and an explicit rank constraint. Low rank representation, similarity learning and a cluster structure constraint are incorporated into the same frame and are applied to medical CT image segmentation. The method comprises the steps that (1) a CT original image is preprocessed; the gray scale is used to correct and enhance the image; histogram equalization is carried out; and a median filter method is used to filter a target area to reduce interference and noise in the image; (2) an algorithm based on correlation matrix self-learning and the explicit rank constraint is used to solve the eigenvalue projection matrix of the image; and (3) a traditional clustering algorithm is used to cluster and split a projection matrix, and a region of interest is calibrated.

The invention discloses a human face set matching method and system based on similarity learning. Part of samples are selected as training samples to carry out training, classifiers are selected, and therefore it is avoided that all samples serve as the training samples to be trained, dimensional reduction processing is carried out on training set samples so that dimensional reduction training samples are obtained, the calculation complexity is prevented from being increased from high dimensional data, the training period is shortened, the training process is simplified, the complex process is avoided, and the training speed is improved. In addition, according to the scheme, the geometric mean of each type of samples in the training set samples is selected to construct a plurality of different classifiers, and the effect of the accurate result through the simple operating process is achieved.

The invention relates to the field of intelligent logistics distribution, and particularly relates to an intelligent culture gene logistics distribution method based on similarity learning. The methodcomprises the following steps of uploading data such as cargo data, customer coordinates, road states and the like to a database; initializing a population; checking whether stopping conditions are met or not; performing roulette selection on two individuals in the population to perform crossing to generate offspring; calculating the similarity between the current individuals; if the similarity between the individuals and existing individuals in the population is lower than a threshold value, performing heuristic local searching on the individuals; and sorting the individuals in the population, thereby selecting out the optimal individuals. A similarity concept is adopted, so that the diversity of the population is controlled to a certain degree in an iteration process of an algorithm, the situation that the population is subject to premature convergence and falls into local optimum is avoided, the searching range of the population is greatly expanded, computing resources can be usedfor more potential solutions as more as possible through the population, and the optimization capability of the population is improved.

The invention provides an image similarity measurement method based on kernel preserving. The method comprises the steps that firstly a loss function is defined, then the loss function is simplified by defining a kernel function, a regularization function is added into the loss function to obtain a complete objective function, finally, the objective function is optimized and solved to obtain a final reconstruction transformation matrix, and two indexes including the accuracy rate and the regularization mutual information amount are adopted to measure the performance of the final reconstructiontransformation matrix. According to the method, the kernel function is defined to minimize a reconstruction error, and the similarity learning is carried out on original image data samples, so that better global relationships among the image data samples are reserved, and based on the similarity information, more accurate clustering is carried out on the images by using a spectral clustering algorithm. The method has universality and can be used for clustering, classifying, recommending systems and other problems, an effective basic module is provided for a method based on similarity learning, and meanwhile great potential is achieved in the application of image mapping to low dimension.

The invention discloses a deep learning remote sensing image registration method based on sub-image matching and application. The method comprises a sub-image matching stage and a transformation parameter estimation stage which are completed by the convolutional neural network. Comprising the following steps: 1, cutting a series of sub-images containing a plurality of features from an image; extracting sub-image features through a sub-image similarity learning network ScoreCNN with a feature vector inner product structure; estimating the similarity of the sub-images in the fusion stage; and according to the similarity, searching matched sub-images with high confidence by using a rapid screening algorithm; and 2, inputting the matched sub-images to the corresponding coordinates in the original image into a transformation parameter estimation network ETPN with a weight structure and position codes, and outputting a transformation matrix between the images to be registered. According to the invention, the problem of algorithm failure caused by insufficient correctly matched features in image registration with large feature change in a traditional registration frame is solved, and meanwhile, the precision of the deep learning registration method based on parameter regression is improved.

The invention discloses a robust measurement based handwriting identification method and system. The method comprises the steps of: constructing a weighted similar map by performing similarity learning on a handwriting training sample; and keeping local features of all training samples while compacting a local intra-class divergence and separating a local inter-class divergence. In order to improve robustness of handwriting description, 1-norm measurement is proposed to be applied in a semi-supervised learning model, so as to design a performance robustness handwriting identification method and system, and output a projection matrix P that can be used for handwriting image feature extraction within a sample and outside the sample. Induction on images other than the sample comprises the steps of: projecting a test sample to the projection matrix P to input an extracted feature into an effective label propagation classifier for classification; and selecting a place of maximum probability in a corresponding category of soft label to determine a category of the test sample, so as to obtain a most accurate character recognition result. Meanwhile, by establishing a ratio model, model parameters are reduced, and the projection matrix P meets the orthogonal property.

The invention discloses an image similarity learning method for extracting image multi-resolution features. The method comprises the following steps: 1, acquiring a chip slot image by adopting a laboratory industrial camera, performing normalization processing on images, combining every two similar or dissimilar single images into a dual-channel image according to similarity and dissimilarity ofthe images through human visual judgment and a hash algorithm to form an input image pair data set, and dividing the input image pair data set into a training set and a test set; 2, constructing a network model, selecting a deep learning framework, specifying a network training target function and an optimizer, and performing similarity learning; 3, performing network model training and testing. The problems that in existing image similarity learning, image feature information extracted through a network is single, image features cannot be accurately described, network training is limited by the data size, and overfitting is likely to happen are solved.

The invention belongs to the field of target re-identification, and provides a target re-identification method and system based on non-supervised pyramid similarity learning. The target re-identification method based on non-supervision pyramid similarity learning comprises the steps of obtaining a to-be-queried sample image and a target scene domain image; and outputting a target image matched with the to-be-queried sample image in the target scene domain through the target re-identification model, wherein a training and updating process of the target re-identification model includes the following steps: carrying out non-supervision multi-scale horizontal pyramid similarity learning on images of a source scene domain and a target scene domain; and automatically labeling the target scene domain sample image according to the similarity and screening out a training sample to train and update the initial model to obtain a target re-identification model. Through continuous iterative training and updating, the model is more and more adaptive to sample data in a target scene domain, and the accuracy of pedestrian target re-identification can be improved.

The invention provides a visual similarity learning method based on depth unsupervised learning, Its main contents include: Compact clustering and batch generation, Convolution Neural Network (CNN) training, local time pooling and multi-instance learning, The process is, The associated initial sample set is first obtained from the sample, the single cost function is then optimized to obtain a compact cluster (a compact and similar sample set with compact location distribution), Then, the CNN is trained alternately, and the generated similarities are subjected to local time pooling operation, and the clusters and batches are recalculated using the obtained similarities, and the similarities among the samples are obtained after many iterations. The present invention solves the problem that the prior visual similarity learning method requires a large amount of manual labeling data and has high calculation cost, can provide more fine similar structure, and has good performance in attitudeanalysis task and classification problem.

A system and method for distributed similarity learning for high-dimensional image features are described. A set of data features is accessed. Subspaces from a space formed by the set of data features are determined using a set of projection matrices. Each subspace has a dimension lower than a dimension of the set of data features. Similarity functions are computed for the subspaces. Each similarity function is based on the dimension of the corresponding subspace. A linear combination of the similarity functions is performed to determine a similarity function for the set of data features.