684results about How to "Improve retrieval accuracy" patented technology

The invention relates to a three-dimensional model retrieval method based on multi-feature relevance feedback, comprising the following steps: a server processes each three-dimensional model in a three-dimensional model database to obtain color view arrays; three-dimensional model features are acquired and synthesized to generate a feature database; features of a two-dimensional sketch offered by a computer client are calculated and matched with feathers in the feature database, the distance between the two-dimensional sketch and each three-dimensional model is calculated, the three-dimensional models as a retrieval result are sorted and outputted according to the distance values; the client labels each retrieval result with 'relevance' or 'irrelevance', returns labeled three-dimensional model information to the server, the server learns the information, classifies the three-dimensional database by SVM fusion method, sorts and outputs the three-dimensional models as a retrieval result; the above steps are repeated until a satisfactory three-dimensional model retrieval result is outputted to a user.

The invention discloses a picture index construction based on keywords and contents and corresponding searching method, wherein the index construction comprises: constructing the keywords inverted index based on keywords according to the picture describing explanation and the name describing explanation thereof; constructing the picture characteristic index based on content by extracting the picture characteristic vector. The searching method comprises: searching the keywords and the keywords set to perform mode matching to obtain a semantic-related picture set based on the searching request submitted by the user; searching in the picture characteristic vector index by the Hash method with characteristic vector sensitive on the applied position for sampling the picture characteristic to obtain a similar picture characteristic vector; returning the picture with high similarity based on the integrative result to the user. The invention can compromise the keyword index and the picture index, not only can use the keywords to improve the searching speed, but also can use the picture index to improve the searching result correlation degree, so that the precision ratio can be improved. The technical scheme can be applied to the searching field of the picture image.

The present invention provides a supervised depth hashing fast picture retrieval method and system. The method comprises: constructing a depth convolution neural network H" for fast image retrieval; after the pictures in the library are sequentially input into the depth convolution neural network H", obtaining real value features, obtaining hash codes after the quantization operation and storing the codes locally; and inputting each query picture q into the depth convolution neural network H", quantifying the picture to obtain the hash code h (q), calculating the Hamming distance between the hash code h (q) and all hash codes stored locally, taking the small Hamming distance as that the similarity is high, sorting the pictures by taking the order, and finally returning the corresponding number of pictures with the highest similarity according to the requirement of the retrieval number. According to the method and system provided by the present invention, based on the existing depth neural network, the learning of the picture feature expression is carried out by using the triple tag data, and a triple quantization loss function is used to construct the supervised depth hashing model, so that fast and accurate image retrieval can be realized.

The invention discloses an antagonistic cross-media search method based on limited text space. The method comprises the steps that a characteristic extraction network, a characteristic mapping networkand a modality classifier are designed, the limited text space is obtained by learning, image and text characteristics suitable for the cross-media search are extracted, and the mapping of the imagecharacteristic from image space to text space is achieved; the difference among characteristic distributions of different modality data is constantly reduced in the learning process by an antagonistictraining mechanism; and thereby the cross-media search is achieved. The antagonistic cross-media search method based on the limited text space has the advantages that behavioral expressions of peoplein cross-media search tasks can be better fitted; the image and text characteristics more suitable for the cross-media search tasks are obtained, and the defects of pre-training characteristics in anexpressive ability are made up; and the antagonistic learning mechanism is introduced, so that the search accuracy is further improved through the maximum minimum game between the modality classifierand the characteristic mapping network.

The invention relates to a cross-media retrieval method based on deep learning and consistent expression spatial learning. By starting with two methods including feature selection and the similarity estimation of two highly-isomerous feature spaces, the invention puts forward the cross-media retrieval method capable of improving multimedia retrieval accuracy to a large extent by aiming at the cross-media information of two modalities including an images and a text. The method disclosed by the invention is a multimedia information mutual retrieval method which aims at two modalities including the image and the text, and cross-media retrieval accuracy is improved to a large extent. In a model which is put forward by the invention, a regulated vector inner product is adopted as a similarity metric algorithm, the directions of the feature vectors of two different modalities are considered, the influence of an index dimension is eliminated after centralization is carried out, an average value of elements is subtracted from each element in the vectors, the correlation of the two vectors subjected to average value removal is calculated, and accurate similarity can be obtained through calculation.

The invention discloses a fast multi-label picture retrieval system and a realization method. The method comprises the following steps: deploying an RPN (Region Proposal Network) for extracting region proposals in a convolutional neural network, extracting region proposal information of pictures, and performing ROI pooling calculation on the region proposal information; after pooling, building a multi-label classification loss function through a fully connected layer according to multi-label information to train the convolutional neural network, and building a weighted three-dimensional loss function to train the convolutional neural network; extracting the hash code of each picture from a picture candidate set through the convolutional neural network after multi-task learning, saving the hash codes to a database, and comparing the hash codes with the hash codes in the database, thus completing picture retrieval. The whole network is trained through multi-task learning of classification and hashing, and therefore, the accuracy of retrieval is ensured. Moreover, the similarity is measured using Hamming distance in the process of retrieval, and the efficiency of retrieval is improved greatly.

The invention relates to the technical field of information retrieval, in particular a semantic relationship network-based cross-mode information retrieval method. In the method, cross-mode association knowledge is acquired by webpage vision spatial analysis, multimedia search engine label relationship analysis, DeepWe interface mode analysis, analysis on the association of data in different modes in composite multimedia, utilization of direct and potential feedback information of users and association reasoning, and a cross-mode association network is constructed; multimode data sets having the same semanteme and different finenesses are acquired by using the acquired cross-mode association knowledge and hierarchical fuzzy clustering; and typical vectors in different modes are selected from each SC, corresponding semantic vector packets are built, and mapping relations are built among the SCs, the typical vectors and the corresponding semantic vector packets. The method can reduce possible errors in each channel, improve retrieval accuracy effectively, support cross-mode retrievals with semantemes of different finenesses defined by users, and support the retrieval by using multimode data files as samples at the same time.

The invention discloses a retrieval system for accomplishing data rapid search and a method thereof; wherein, the system comprises: a retrieval document establishing unit, a retrieval document storing unit and a retrieval serving unit; wherein; after forming a main document storage based on initial data source and increased data source, the retrieval document establishing unit establishes more than one retrieval documents with different levels by using different attributes of data in main document according to specific retrieval service requirement, and stores the retrieval documents in the corresponding retrieval documents document storing unit, and supplies the retrieval documents to the retrieval serving unit for the retrieval documents. The system and method of the invention accomplish rapid retrieval of increased data source with high retrieval efficiency and retrieval precision, so as to greatly enhance user experiencing in retrieval service.

The invention discloses a personalized search method for Web service recommendation. The personalized search method comprises the following steps of: 1, preprocessing a WSDL (Web Services Description Language) file, i.e., forming a bag of words through two preprocessing steps of removing stop words and extracting stems; 2, extracting user interest, i.e., calculating weight of each word in the bag of words by using an improved TF-IDF (Term Frequency-Inverse Document Frequency) formula, and multiplying by a time decay factor of the word to obtain a new weight; selecting previous k words according to the weight from large to small as interest words of a user and corresponding weight of each word to form a k-dimension user interest vector; 3, calculating interest similarity, i.e., setting a similarity threshold and selecting the users with interest similarity exceeding the threshold as neighbor users of a target user; and 4, ordering service search results, calculating a recommended predicted value of the service according to similarity of neighbor users and the frequency of selecting service of the users, and arranging the searched results in a descending order according to the recommended predicted value, thereby obtaining the personalized search result.

The invention discloses a latent semantic image retrieval method of region-oriented multi-feature integration and multi-level feedback. It uses result list returned by the initial keyword search, extracting a variety of region-oriented images characteristics, constructing attribute-image matrix, using latent semantic indexing algorithm to get the semantic space of image sets and semantic features of each image, and then using similar images by users feedback to construct or update image query vector, searching again the semantic space, calculating image semantics features and images inquiries vector similarity, getting outcome sets by descending order, and repeatable retrieval. The invention takes full advantage of image content information, making up for the deficiencies of the keyword search, and through the region-oriented multi-feature integration, enhances image content information from the bottom physical layer to the object layer, then further enhances to the semantic layer by HCI feedback, thereby reducing the gap between the image bottom features and high-level semantic, and allowing Web image retrieval to get higher retrieval accuracy.

The invention relates to a garment image retrieval method based on segmentation and feature matching, and belongs to the field of computer vision and image application. The method comprises the steps that firstly, a garment image to be detected is input, and the garment image is segmented by extracting HOG feature information of the garment image to be detected and HOG feature information of a small image library for assisting in segmentation; secondly, the color feature and Bundled feature of the garment area of the segmented image to be detected are extracted, and feature extraction of a garment to be detected is achieved; thirdly, the feature similarity between the image to be detected and a large image library is calculated according to the garment feature of the image to be detected and a feature library obtained by preprocessing the large image library for retrieval, and feature matching is carried out; finally, the large image library is searched for garment images similar to the image to be detected according to the feature matching result, and search results are output according to the similarity sequence. The retrieval method has high accuracy.

The invention discloses a three-dimensional model search method based on mesh segmentation. The method comprises the following step of analyzing and constructing a segmentation field through a hierarchical spectrum, and particularly comprises the steps of judgment of concave vertexes, constructing of a Laplacian matrix, matrix decomposition, selection of low-frequency feature vectors, generation of sub feature vectors, weight calculation of the sub feature vectors, and constructing of an edge symbol matrix. Contour lines are sampled in the segmentation field and are grouped and merged through a grouping-merging algorithm to obtain a plurality of candidate contour sets, the final segmentation boundary is determined according to the weight of each contour line in the candidate contour sets, and three-dimensional models are automatically segmented. Three-dimensional model mixing feature description sub-matrixes are obtained by calculating the feature description sub-matrix of each segmentation block of the three-dimensional models, the similarity of the mixing feature description sub-matrixes of each three-dimensional model in a three-dimensional model database to be searched and a target three-dimensional model to be searched is calculated, the similarity values of the three-dimensional models are ranked and output from low to high, and the three-dimensional model searching is achieved.

The invention discloses a cross-modal image text retrieval method of a hybrid fusion model. The method comprises the following steps of: in an early fusion structure, firstly combining local visual area features and original global features of a text to obtain a unified cross-modal fusion representation, and then taking the fusion features as input, enhancing interaction between the local visual features and the language information in a subsequent embedded network; meanwhile, on the basis of a traditional late fusion structure, inputting an original image and sentence features into a vision encoder and a text encoder respectively for intra-modal feature enhancement, and enriching semantic information of respective modals; and finally, the whole network similarity is a weighted linear combination of early fusion similarity and late fusion similarity so that complementation of early fusion in a cross-modal learning layer and late fusion in an intra-modal learning layer is realized, and potential alignment between image and text modals is completed.

The invention discloses a deep hash and GPU acceleration-based large-scale image retrieval method. According to the method, classification loss functions and comparison loss functions are combined onthe basis of a hash method for picture pairs by adoption of a multitask deep learning mechanism, and in the quantification process, similarities between the picture pairs are kept, semantic information of pictures is kept as much as possible, and classification tasks and quantification tasks carry out mutual guiding and learning; and meanwhile, a full-connection layer of a quantification network is replaced by a local connection module, so that redundant information between features is decreased. According to the method, deeper networks are designed and realized, and the deeper networks can obtain feature expressions in general. On the basis of Han/Ming sorting, a GPU-based multilayer parallel retrieval method is realized. The method is capable of improving the retrieval precision and achieving an effect of delaying the retrieval of a single image in a million-scale image library to be 0.8ms.

The invention discloses an image retrieval method based on multiple intelligent algorithms and an image fusion technology. According to the method, on the basis of a reinforced learning and genetic algorithm and a clustering algorithm, an intelligent learning framework for active learning is constructed; the image fusion technology and the genetic algorithm are used during relevant feedback, so that an inquiry vector and a similarity matching model are corrected; and the inquiry precision and the inquiry efficiency are improved. The image retrieval method has the advantages of higher inquiry precision and inquiry efficiency and relatively high robustness for translation, rotation and scale transformation; furthermore, after a certain number of inquiries and learning, intelligent retrieval can be realized; and the inquiry precision and the inquiry efficiency are further improved.

The invention discloses a vertical engine searching method for domain body restraint, which comprises the following steps of: establishing a domain body library and constructing a domain body generator and a domain body importer; performing semantic analysis and body description on network resources by using a web crawler with a semantic analysis filter from various body models of the domain body library, and automatically performing resource conformity calculation and classification to form classification information with semantic features; establishing a semantic relation between the network resources, forming domain resources with semantic relation, and finishing body description of single network resource and storing the network resources; and performing semantic rewriting and mapping on retrieval according to the body, finishing semantic analysis and expansion of the retrieval, taking the body-based resource and retrieval as input, and finishing expansion of the body library and rule restraint expansion through rule learning and modes so as to form a secondary body of the domain body. The method has the advantages that the method with semantic restraint saves the retrieval time, improves the retrieval precision and realizes the advantage of semantic supporting retrieval.

The invention discloses a method for detecting similar texts on the basis of text picture retrieval. The method includes steps of establishing document libraries; establishing text picture libraries; extracting features of pictures in the text picture libraries and reducing dimensions; segmenting retrieval documents to obtain retrieval picture sets; extracting features of pictures in the retrieval picture sets and reducing dimensions; measuring cosine similarity of the retrieval picture sets; filtering full-test similarity of retrieval results; outputting the retrieval results. The method has the advantages that diversified multilayer convolutional neural network modules are integrated with one another for training CNN (convolutional neural network) feature description operators, and accordingly text images can be deeply visually represented; the dimensions are reduced by means of PCA (principal component analysis) compression, and accordingly the similarity measurement efficiency can be improved; full-text similarity filter models are built from the filter aspect of the retrieval results, improvement can be carried out, accordingly, the similarity of the retrieval results can be updated, the retrieval precision can be improved, optional multiple character texts can be directly recommended and retrieved, and the method is excellent in similar text detection capacity and can be used for checking the repeatability of the texts or recommending similar literature.

The invention discloses a Hash image retrieval method based on deep learning and low-rank matrix optimization, and the method comprises the following steps: S1, obtaining image data, carrying out themarking and preprocessing of the data, constructing a data set of image retrieval, and dividing the data set into a training set and a test set; S2, constructing a deep feature extraction network, andconstructing a deep Hash network trunk; S3, inputting the training set into a deep Hash network trunk, and constructing a Hash network based on a maximum probability likelihood and a low-rank regularization loss function; S4, training the Hash network; S5, respectively inputting the test set image and the training set image into a Hash network, generating a binary Hash code, and calculating a mutual Hamming distance; and S6, returning a picture with the minimum Hamming distance in the training set as a retrieval result. According to the image retrieval method based on Hash representation, theproblems of ring breakage of similarity information and large quantization error caused by the fact that binary continuous value features are directly coded into a Hamming space are solved, and the performance of the image retrieval method based on Hash representation is improved.

The invention provides an information retrieval method based on a spurious correlation feedback model. The information retrieval method based on the spurious correlation feedback model comprises the following steps: fusing word correlation to a spurious correlation feedback model to realize information retrieval, respectively generating a query expansion word which uses the importance degree of a candidate expansion word as a feature and a query expansion word which uses the relevancy of the candidate expansion word and the query expansion word as a feature when the query expansion word is generated in a spurious correlation document set, and then binding the two query expansion words into the original query expansion word, and finishing final information retrieval; and when generating a query expansion word which uses the relevancy of the candidate expansion word and the query expansion word as a feature, calculating the relevancy between a query word and a candidate word which appear at different positions in a document by using a kernel function. By the method, the distribution condition of the query word and the candidate word can be highlighted, the candidate word which has higher degree of correlation on query thematic words is selected, and therefore, the accurate candidate word is positioned and precision of expansion query and final retrieval is improved due to additional relevancy information.