80results about How to "Enhanced Semantic Information" patented technology

The invention relates to the technical field of traffic image processing, and discloses a driving scene target detection method based on deep learning and multi-layer feature fusion. The method comprises the following steps of 1) collecting a video image to serve as a training data set, and performing preprocessing; 2) building a training network; 3) initializing the training network to obtain a pre-training model; 4) performing training on the training data set by using the pre-training model obtained in the step 3), thereby obtaining a training model; 5) collecting a front image by using a vehicle-mounted camera, and inputting the image into the training model obtained in the step 4), thereby obtaining a detection result. The multi-layer feature fusion method based on a feature pyramid is adopted to enhance semantic information of a low-layer feature graph, so that the feature extraction quality of the network is improved, and higher detection precision is obtained.

The invention belongs to the technical field of information retrieval, and provides a full-text retrieval system based on two-level semantic analysis. The system comprises a query information receiving module, a sentence summary level semantic processing module, a retrieval module, an index library, an index module, a text level semantic processing module, a result set processing module and a data server. The system performs summary processing and repeat processing for a query statement to repeat the query statement submitted by a user from sentence summary level semanteme, and latent semantic information and thematic information are extracted in the text level to eliminate retrieval result documents not conforming to text level semanteme, so that the full-text retrieval system achieves two-level semantic analysis in the 'sentence summary level' and the 'text level'. By improving semantic processing capacity of query statements and text sets, the full-text retrieval system is more human-friendly and intelligent, and provides more convenient and accurate services for users.

The invention provides a Chinese short text subjective question automatic scoring method using an LSTM neural network. The method comprises the steps that firstly, an answer text is segmented, and thetext is converted into a word sequence; secondly, a vectorization expression of each word in the answer text is obtained, and an answer text mapping matrix is constructed; thirdly, the LSTM neural network is used for carrying out operation on the answer text mapping matrix, output of all or a part of hidden layers is obtained to obtain a semantic feature matrix of the answer text; fourthly, down-sampling is conducted on the semantic feature matrix by utilizing a pooling algorithm to obtain a semantic feature vector of the answer text; fifthly, the semantic feature vector of the answer text isgiven to a classifier, and the category of the answer text is predicted; sixthly, the many-to-one relationship between the category where the answer text belongs and the score is considered, and thescore of the answer text is determined according to preset mapping between the category and the score. According to the method, answer text semantic information can be effectively mined without depending on subjective question standard answers, and Chinese short text subjective question automatic scoring is achieved.

The invention belongs to the information retrieval technology and provides a full-text retrieval system based on dynamic semantic analysis. The full-text retrieval system comprises a query information receiving module, a query statement semantic processing module, a semantic knowledge base, a word contribution degree dynamic calculation module, a retrieval module, an index database, an index module, a result processing module and a document set. According to the full-text retrieval system based on dynamic semantic analysis, word contribution degree information can be dynamically calculated and updated, word sense disambiguation can be conducted on query expansion statements, the query expansion statements meeting semantics are screened out, and full-text information retrieval supporting dynamic semantic analysis is achieved. The full-text retrieval system has the advantages of being higher in precision ratio, recall ratio and dynamic performance.

The embodiment of the invention provides a semantic segmentation method, a semantic segmentation device, a semantic segmentation system and a storage medium. The method comprises the steps of: obtaining an image to be processed; and inputting the image to be processed into a U-shaped network, so that a semantic segmentation result of the image to be processed output by the U-shaped network is obtained, wherein the contraction path of the U-shaped network comprises n convolutional modules, which are connected in sequence; output features of the ith convolutional module in the n convolutional modules are combined with output features of at least one convolutional module after the ith convolutional module; the combined features are in jump connection with the output end of a deconvolution layer in the contraction path of the U-shaped network and corresponding to the ith convolutional module, wherein n is an integer, which is greater than 1; and i is greater than or equal to 1 and less than n. According to the semantic segmentation method, the semantic segmentation device, the semantic segmentation system and the storage medium in the embodiment of the invention, because of adoption ofshallow features and deep features, the U-shaped network can be well improved in a fusion manner; and thus, a relatively accurate semantic segmentation result can be obtained.

The invention discloses a rule and path combined knowledge graph representation learning method, which comprises the following steps of: firstly, extracting a logic rule from a knowledge graph, and encoding and representing the logic rule; then, on the basis of the rule subjected to encoding representation, completing a relation semantic combination operation in the relation path and establishingsemantic association between the relation pairs; and finally, in combination with the triplet, the relation path vector representation between the entities and the semantic association constraint between the relation vectors, jointly constructing an energy equation, and obtaining a minimum evaluation function. The method has the advantages that not only is the accuracy of relation representation improved, but also semantic association among the relations is established by the aid of rules, vector representation of the relations with the semantic association is constrained, more semantic information is added into the vector representation of the relations, and the precision of the vector representation of the relations is improved.

The invention discloses a deep learning problem classification method combining a multi-level attention mechanism, and relates to the field of information processing technology. The deep learning problem classification method comprises the following steps: constructing an interrogative word vector set, wherein the interrogative word vector set includes interrogative word vectors and common word vectors including interrogative word information; extracting window mapping of interrogative sentences by convolution operation according to the interrogative word vector set; extracting sequential characteristics of the interrogative sentences according to the window mapping; and classifying the interrogative sentences according to the sequential characteristics. Through adoption of the method, semantic information of interrogative words in the interrogative sentences is enhanced. Moreover, a convolutional neural network and a long-term and short-term memory model are fused through the attention mechanism in deep learning, thereby effectively increasing the accuracy of problem classification.

The invention discloses a convolutional neural network cloth defect detection method based on an extreme learning machine. The method comprises the steps of carrying out feature extraction by constructing a convolutional neural network; fusing the extracted features; extracting cloth defect candidate boxes on the fused feature layer; in a cloth defect detection stage, carrying out regression on the extracted cloth defect candidate boxes by using the convolutional neural network; classifying the cloth defect candidate boxes by using an extreme learning machine; calculating loss by combining classification and regression results with real labels of sample pictures; updating the weight in the network by using a random gradient descent method based on the obtained loss; carrying out continuousiterative training until the loss of the network converges to a minimum value or reaches a preset training round number, and thus obtaining a trained cloth defect detection network model based on theextreme learning machine, so cloth defect detection can be carried out. The overall performance of cloth defect detection is effectively improved, and the cloth defect detection method has higher cloth defect detection accuracy.

The invention relates to an object detection method based on semantic segmentation enhancement, which comprises the following steps: preparing a labeled image; image set partition; deep convolution neural network structure based on semantic segmentation enhancement is designed for object detection. The deep convolution neural network based on semantic segmentation enhancement includes three main parts: backbone subnetwork, segmentation subnetwork and detection subnetwork. The backbone subnetwork is used to extract the general feature of the image, which is class-independent feature. The segmentation subnetwork extracts the features of semantic segmentation based on the backbone subnetwork, and predicts the segmentation heat map of each class of objects. The thermal map of each class of objects is used as a priori knowledge of the class, and is fused with the features extracted from the detection subnetwork to generate the class-related features. Each class of objects has the features of the corresponding class, and the features reflect the characteristics of the class of objects significantly. Model training.

An attention remote sensing image target detection method based on an anchor frame comprises the following steps: S100, embedding a jump connection feature pyramid module and a position attention module based on the anchor frame into a target detection model, and establishing an end-to-end single-stage remote sensing target detection model; S200, detecting a remote sensing target by adopting the end-to-end single-stage remote sensing target detection model; S300, outputting a detection result of the remote sensing target. The method can improve the precision and speed of the target detection model.

The invention discloses a three-dimensional building information model construction and automatic updating method based on generalized point cloud, relates to the technical field of building intelligent management, enhances the safety analysis capability of a house management department and the self-control capability of house price evaluation of real estate providers, and provides an effective tool for decision making and participation in a decision making process for governments and residents. According to the specific scheme, the method comprises the following steps: S1, creating a three-dimensional building information model based on semantic segmentation and information enhancement; and S2, automatically updating the three-dimensional building information model. A laser scanning technology and an oblique photography technology are used for researching fusion understanding object semantics of a laser point cloud and an oblique image, firstly, feature extraction is conducted on thelaser point cloud and the oblique image, the incidence relation of the point cloud and the image under geometric and radiation measurement space is established, a feature matching technology under structural relation constraints is used, and alternate energy transmission is realized.

The invention discloses an image multi-tag marking algorithm based on multi-example package feature learning, and the algorithm comprises the steps: obtaining a set of image blocks of all training images; extracting the features of a color histogram and the features of a direction gradient histogram of each image block of the set of the training images; enabling one training image to serve as an image package, and obtaining an image package structure needed by a multi-example learning framework; enabling the examples in all image packages in the set to form a projection example set, enabling each image package to be projected towards the projection example set, and obtaining the projection features of the image packages; selecting the features with the high discrimination performance as the classification features of the image packages; importing the classification features of the image packages of the learned training image set into an SVM classifier for training, obtaining the parameters of a training model, and predicting a test image tag through employing a trained SVM classifier. The algorithm is simple in implementation, and a trainer is mature and reliable. The algorithm is quick in prediction, and achieves multiple image tags better.

The invention discloses a deep learning small target detection method and device based on cascade fusion and an attention mechanism, and the method comprises the following steps: S1, inputting a to-be-detected image, and carrying out the preprocessing of the to-be-detected image; s2, performing feature extraction on the preprocessed image by using a deep convolutional neural network based on cascade fusion and an attention mechanism, extracting to obtain target image features, performing feature fusion by a feature cascade fusion layer based on a cascade feature fusion structure, enabling the spatial attention mechanism layer to obtain a semantic mask of the small target area, fusing the semantic mask with the original features channel by channel, and outputing extracted target image features; and S3, carrying out prediction and post-processing on the extracted target image features to obtain a final target detection result, and outputting the final target detection result. The invention can achieve small target detection based on deep learning, and has the advantages of being simple in implementation method, low in cost, high in detection efficiency and precision, flexible in operation and the like.

The invention discloses an educational resource semantic annotation method. For different learning resources, resource information is extracted and a learning resource base is constructed, mapping of labels inside a learning resource library and a label library is achieved in a mapping mode, corresponding labelling information is extracted to be stored inside a label information base, and accordingly multi-label labelling of the learning resources is achieved. According to the educational resource semantic annotation method, experience of teachers and student grading are comprehensively taken into consideration, the resources and labels are labeled in a one-to-many mode, the corresponding semantic information is increased for the resources, and a user can accurately search the required resources.

The invention belongs to the field of computer application, and discloses a method for perfecting a medical term knowledge base. The method comprises the steps of extracting any term from a medical term set to serve as a candidate term, and then extracting any term from a medical term knowledge base to form a term pair set; extracting any term pair in the term pair set, and calculating the correlation of the term pairs; and sorting the correlations of all term pairs in the term pair set, and fusing the term pairs as synonyms into the medical term knowledge base according to a correlation sorting result. Experiments show that by adopting the embodiment of the invention, term alignment can be accurately carried out, and a knowledge base can be effectively perfected.

The invention provides a cross-language document similarity detection method, and belongs to the technical field of document similarity comparison. The method comprises the following steps: firstly, respectively converting a source document and a target document to be compared into intermediate documents based on the same language word; searching similar intermediate representation sets between the two intermediate documents so as to establish a mapping set; and finally finding out similar text sections between the source document and the target document through the mapping set according to amethod of searching the similar text sections. The cross-language document similarity detection method provided by the invention has the beneficial effects that the difficulty in cross-language document similarity detection can be better solved and better detection results can be obtained.

The present invention provides a semantic Web service automatic combination method based on a body. The method comprises the steps of obtaining a Web request service; searching all Web services (WS) of which the exit semantics match the exit semantic of the Web request service from a Web service list; using a from-right-to-left matching combination algorithm, searching all WS' of which the exit semantics match the entrance semantics of the WS from the Web service list, and combining the WS and the WS' into the new WS until all entrance semantics of the Web request service match the WS. The semantic Web service automatic combination method based on the body of the present invention solves the problems in the prior art that the Web service combination is limited by the actual application scenes, and the combination quality is low.

The invention belongs to the technical field of big data retrieval, and provides a multi-modal retrieval method and system based on weak supervision hash learning. In order to solve the problem of incomplete pairing information among modals, the method comprises the following steps: acquiring a to-be-retrieved sample, and carrying out hash code calculation on the to-be-retrieved sample; performing 0/1 XOR operation on the hash code of the to-be-retrieved sample and the hash code in the retrieval database to calculate a Hamming distance, and returning similar data from small to large according to the Hamming distance; the construction process of the retrieval database comprises the following steps: establishing a semi-supervised and semi-paired cross-modal hash target function based on intra-modal pairing similarity, inter-modal pairing similarity and complemented label information of each modal; hash representation is obtained by optimizing an objective function of semi-supervised and semi-paired cross-modal Hash, sampling is carried out from the Hash representation, then corresponding partial cross-modal similarity information is embedded into Hash function learning, and finally, a retrieval database is generated by utilizing the embedded Hash function. According to the method, the calculation complexity is reduced, and the retrieval precision is improved.

The invention relates to the field of logistics detection, and discloses a method, a device and equipment for identifying a uniform sign, and a storage medium, which are used for improving the identification accuracy of the uniform sign. The method for identifying the uniform sign comprises the steps: acquiring original image data to be processed, performing image enhancement processing and size conversion processing on the original image data to obtain a target image, wherein the original image data are employee image data in a preset scene; through a preset target model, carrying out featuremap extraction processing and bounding box generation processing on the target image to obtain initial region information; performing portrait detection processing and region segmentation processingon the initial region information to obtain candidate region information; performing convolution processing and anchor box matching processing on the candidate region information through a target model to obtain target region information; and recognizing the image information in the target area information according to preset uniform sign information to obtain a uniform recognition result.