43results about How to "Enhanced representation" patented technology

The invention discloses a deep learning super-resolution reconstruction method based on residual sub-images; residual sub-images are effectively combined with deep learning method based on convolutional neural network, super-resolution reconstructed images are clearer, and reconstruction speed is higher. By increasing the depth of convolutional neural network, a network model acquired by learning has higher nonlinear expression capacity and image reconstructing capacity; in addition, by introducing residual sub-image process, preprocessing based on traditional interpolation algorithm is removed, and fuzzy effect due to the interpolation algorithm is avoided. By making ingenious use of residual sub-images, it is possible to transfer deep learning convolutional operation from high-resolution space to low-resolution space, and accordingly reconstruction efficiency of super-resolution algorithm is increased at the premise of improving super-resolution reconstruction effect.

The invention provides a real image blind denoising method based on a deep residual network. According to the method, an RGB spatial clear image set is selected through an image dataset, and an RGB spatial image group set is constructed through spatial transformation; images under multiple scenes are shot through multiple cameras, and real image groups are constructed according to real clear images and real noisy images shot by each camera under each scene, and a real image group set is constructed; multiple RGB spatial image groups in the RGB spatial image group set and multiple real image groups in the real image group set are randomly selected to construct an image training set, and a preprocessed image training set is obtained through preprocessing; remaining RGB spatial image groups in the RGB spatial image group set and remaining real image groups in the real image group set are used to construct an image test set; and the preprocessed image training set is used as input to construct an image denoising residual convolutional neural network, the neural network is trained in combination with residual learning and a batch normalization strategy, and the image test set is denoised. The method has the advantages that convergence speed is high, and the denoising effect is good.

The invention discloses a global and local feature embedding and image-text fusion sentiment analysis method and system. The method comprises the following steps: firstly, extracting global features of an image by using a convolutional neural network, training a target detection model by using transfer learning, detecting and positioning a local region carrying emotion in the image, extracting local region features, and embedding the local region features into the global features to jointly train an image emotion classification model to obtain an emotion polarity probability of the image; expressing the text as a word vector containing rich semantic information, and inputting the word vector into a bidirectional LSTM capable of extracting text semantic context features to perform sentimentclassification to obtain the polarity probability of sentiment of the text; and finally, according to a later fusion formula, obtaining an emotion polarity probability after image and text fusion, and performing image-text emotion analysis. The method can effectively pay attention to the emotion information of the image-text, and improves the accuracy and robustness of image-text emotion classification through the extraction of global and local emotion features of the image and the fusion of text information.

The invention discloses a facial expression identification method and relates to a method for identifying images. The facial expression identification method is a facial expression identification method which extracts expression textural characteristics using a center symmetrical ternary patterns (CSTP) algorithm. The method comprises steps of preprocessing a facial expression image; extracting facial expression textural characteristics on each sub-block of the facial expression image; and determining final facial expression textural characteristics of the facial expression image. Therefore, the facial expression identification is completed. The method overcomes problems that texture description is not fine due to a complex identification background and the identification rate is not high in the prior art.

The embodiment of the invention discloses a face recognition method, device and facility based on a structured analysis dictionary and a computer storage medium. The face recognition method comprisesthe steps of inputting a face training sample set and a face testing sample set into a face recognition model, and performing parameter initialization on the face recognition model, wherein the face recognition model comprises a plurality of sub-classifiers, and each sub-classifier corresponds to a category in the face training sample set; performing structured dictionary learning on each categoryof data sets contained in the face training sample set to obtain a sub-dictionary of each category of data sets and corresponding sub-sparse codes; performing approximate analysis on the sub-sparse codes of each category of data sets to obtain a sub-projection of each sub-sparse code; training each sub-classifier in the face recognition model according to each sub-projection and the face trainingsample set; and performing classification and recognition on the face testing sample set by using each sub-classifier in the face recognition model. The technical scheme of the invention effectivelyimproves the efficiency and accuracy of face recognition.

The invention relates to a device (1) which comprises two parallel partial devices (5, 7) that are provided with three signal inputs (51, 71) each for achieving a redundant two-out-of-three rotational speed detection. Every partial device (5, 7) comprises an arithmetic unit (53, 73) which is used to calculate the rotational speed of the rotating machine part (3). Every partial device (5, 7) furthermore comprises a signal output for outputting the rotational speed calculated by the respective arithmetic unit (51, 71). Matching signal inputs of the partial devices (5, 7) are connected in parallel to the output signal (11) of a rotational speed sensor (9). In an error case, the function of the inventive device can be completely taken over by one of the partial devices (5, 7) so that both a faulty output signal (11) of a rotational speed sensor (9) and a failure of one of the partial devices (5, 7) can be tolerated.

The indoor scene semantic annotation method comprises the following steps: inputting a color picture and a depth picture; entering the neural network, wherein the color picture and the depth picture pass through a conv1 and a conv2 _ x respectively; entering a first attention mechanism module ARF _ 1, and obtaining a feature map through calculation of the ARF _ 1; entering conv3 _ x to carry out convolution calculation; entering a second attention mechanism module ARF _ 2, and obtaining a feature map through calculation of the ARF _ 2; entering conv4 _ x to carry out hole convolution calculation; entering a third attention mechanism module ARF _ 3, and obtaining a feature map through calculation of the ARF _ 3; entering conv5 _ x to carry out hole convolution calculation; entering an attention mechanism mixing module ARMF for calculation; entering a space pyramid module SPP to realize multi-level context information fusion; obtaining a semantic annotation result graph.

The invention discloses a personality detection method based on multi-modal alignment and multi-vector representation. The personality detection method comprises the following steps: resampling voiceand video modal data according to each epoch; inputting the plurality of samples and the text modal data thereof into an intra-modal representation module for independent coding to obtain a voice sequence, a video sequence and a text sequence; inputting the voice sequence, the video sequence and the text sequence into an inter-modal alignment characterization module, and splicing after pairwise alignment and interaction to obtain enhanced voice characterization, video characterization and text characterization; respectively splicing all the voice representations, all the video representationsand all the text representations to obtain a voice vector, a video vector and a text vector, and inputting the voice vector, the video vector and the text vector into a convolutional neural network tobe converted into at least two types of personality vectors; and respectively linearizing the at least two types of personality vectors, and mapping through a sigmoid function to obtain prediction probabilities of the at least two types of personality characteristics. Modal representation is enhanced through pairwise interaction of the three modal data, the discrimination capability of the modelis improved, and a more accurate prediction result is obtained.

The invention discloses a method for solving the evaluation parameter weight of an oil-containing gas region. The method comprises the steps of (1) dividing zone evaluation units, determining the boundaries of the evaluation units, (2) determining the parameter characteristic values of the reservoir forming conditions of the evaluation units, mainly including hydrocarbon source rock conditions, migration conditions, reservoir conditions, trap conditions and storage conditions; (3) obtaining evaluation parameter scores of the reservoir forming conditions by adopting an assignment method according to the evaluation criteria of the evaluation parameters of the reference zone; 4) calculating the accumulation coefficient Q of the evaluation unit; 5) taking the reservoir forming condition coefficient as a constraint to obtain the weight of the evaluation parameter of the oil-gas-containing zone, and 6) performing comprehensive evaluation: multiplying the score value of each reservoir formingcondition of the evaluation unit by the weight of the corresponding parameter of the oil-gas-containing zone, then performing accumulation to obtain the total score value Z of the evaluation unit, performing sorting, and taking the evaluation unit in front of the sorting as a favorable optimization target. According to the method, the parameter weight can be objectively determined under the constraint of the regional geological background, so that the oil-gas-containing zone can be reasonably evaluated, and a favorable exploration target can be found.

The invention discloses a process monitoring method based on simultaneous dimension reduction and dictionary learning. The method comprises two steps of offline dictionary learning and online fault monitoring. At the offline dictionary learning stage, a simultaneous dimension reduction and dictionary learning method is put forward to carry out dictionary learning. At the online fault monitoring stage, three functions of fault detection, mode recognition and fault isolation are realized; fault detection is carried out on test data; if the test data are fault data, fault diagnosis is carried outto determine the fault position; and if the test data are not the fault data, mode recognition is carried out. According to the invention, problems of high-dimensional and multi-modal characteristicsof data are solved simultaneously by using projections and dictionaries learned by a simultaneous dimension reduction and dictionary learning method at the offline learning stage and a SPE statisticsamount is configured for the test data by using a projection matrix at the online fault monitoring stage, thereby reducing the calculation complexity of the online monitoring process and improving the real-time performance of online monitoring. Therefore, more spatial information of original data can be learned and reserved and the expression and identification capability of the dictionary is enhanced.

The invention discloses a method for recommending a next interest point based on a user relationship embedding model, which comprises the following steps of: respectively establishing a friend relationship table and a preference similarity relationship table according to friend relationships and user historical sign-in record data, and establishing a user relationship graph through the two tables;obtaining a user relationship sequence by adopting a random walk algorithm according to the established user relationship diagram, and obtaining a low-latitude embedding vector of each user by the user relationship sequence through a Word2Vec word embedding model; and initializing user embedding layer parameters of the neural network by using the low-latitude embedding vector of the user, and giving a next interest point recommendation by using a gating cycle unit. According to the next interest point recommendation method based on the user relationship embedding model, the preference similarity relationship is introduced into the recommendation model, the representation of the user relationship is enhanced, the limitation problem of the recommendation model caused by only considering theone-sidedness of the friend relationship of the user in the existing method is solved, and the accuracy of the recommendation model is effectively improved.

The invention discloses an improved DCNN music genre classification method. The method comprises the steps of inputting a training set and a verification set, extracting MFCC features of the audio information, generating a frequency spectrum, carrying out frequency spectrum cutting, inputting a network model, training a model, verifying the model, judging whether a specified batch is reached or not, and outputting a model. According to the method, self-adaption of channel dimensions is achieved through a function, so that the coverage range of interaction of local area cross channels is ensured, an ECA module is more effectively integrated into an existing DCNN architecture, obvious performance gain is brought to a network model, and then the working efficiency of music genre classification is improved. Through the Mel-frequency cepstrum coefficient, the perception characteristics of a human auditory system are simulated, and the classification precision is further improved.

The invention relates to a time slot resource expansion graph construction method. The method comprises the following steps: S1, establishing a contact table according to a connection relationship among nodes of a satellite network; S2, dividing time slots, defining a time slot resource expansion graph, obtaining a snapshot graph of the time slot resource expansion graph in each time slot according to the contact table, and determining the types of nodes, links and weights in the time slot resource expansion graph; S3, constructing an adjacent matrix and a feature matrix of a node connection relationship of each time slot snapshot graph; S4, constructing an adjacent matrix and a feature matrix of a node connection relationship between the adjacent time slot snapshots; S5, constructing an adjacent matrix and a feature matrix of the time slot resource expansion graph. The time slot resource expansion graph can directly skip to other time slots through the connection relation between adjacent time slots to search the data transmission path from the source node to the destination node; and the link weight parameters in the graph model are properly added and deleted according to the service transmission service quality, so that the calculation complexity of a routing algorithm can be reduced.

The invention relates to a time sequence classification method, system, medium and device based on multi-representation learning. The method comprises the following steps: (1) carrying out the multi-feature coding of a given time sequence based on different time sequence representation strategies; (2) realizing representation fusion and enhancement by using a residual network and a bidirectional long-short-term memory network; and (3) completing classification by using a multi-layer perceptron network, and realizing classification interpretability by using an attention mechanism. According to the method, a multi-channel time sequence representation learning model is constructed, so that the time sequence characteristics can be comprehensively understood based on various representation strategies. According to the representation fusion model based on the residual network and the bidirectional long and short term memory network, multi-view representation can be effectively fused and representation enhancement can be realized, so that the classification precision is effectively improved. According to the method, the important time sequence characteristics of the time sequence can be effectively recognized on the basis of the attention mechanism, namely, the interpretability basis of the classification result can be provided, and the classification interpretability is realized.

The invention belongs to the technical field of data representation, and provides a cross-modal video time sequence action positioning method and system based on a time sequence-space diagram, and the method comprises the steps: receiving video data and natural language query information; determining a natural language query feature representation based on the natural language query information; determining candidate video time sequence action segment feature representations based on the video data; on the basis of the candidate video time sequence action segment feature representations and the natural language query feature representation, predicting the time sequence offset of the candidate video time sequence action segments and the correlation between the candidate video time sequence action segments and natural language query information; and performing offset correction on the candidate video time sequence action segment with the highest correlation score to obtain a final target video time sequence action segment positioning result. According to the method and the system, the time sequence offset and the correlation score of the corresponding candidate video time sequence action segment are predicted through the two feature representations, and the accuracy of video time sequence action positioning is greatly improved.

The invention discloses an automatic heart sound diagnosis system based on deep learning, and belongs to the field of medical signal processing, and the system comprises: a heart sound signal processing module which is used for preprocessing a heart sound signal and segmenting the preprocessed heart sound signal into a plurality of heart sound signal segments with the same length; and a heart sound detection module which is used for respectively inputting the obtained heart sound signal fragments into a pre-trained heart sound detection model to obtain a heart sound category of each heart sound signal fragment, and synthesizing the heart sound type of each heart sound signal segment to obtain the heart sound type of the to-be-detected heart sound signal. According to the automatic heart sound diagnosis system based on deep learning provided by the invention, the heart sound signals are automatically detected in an end-to-end manner, the multi-level time sequence features extracted by the LSTM and the local features extracted by the CNN are combined, the advantages of the CNN local feature extraction capability and the LSTM long-term dependence capture capability are fully utilized, and the classification precision is greatly improved.

The invention discloses a recommendation method and system based on a knowledge graph and a graph convolutional network. The recommendation method based on the knowledge graph and the graph convolutional network specifically comprises the following steps of obtaining a high-order embedded representation of a user; according to the high-order embedded representation of the user, obtaining an article embedded representation; and predicting the user according to the obtained high-order embedded representation and the article embedded representation of the user. According to the application, user representation can be enhanced, and article features are aggregated and propagated through different relation weights in the knowledge graph, so that remote interests of users are explored.

The invention discloses a content big data-oriented small sample relation extraction method and device. The method comprises the following steps: firstly, performing vectorization processing on sentences by utilizing a pre-training language model and a character-level LSTM; secondly, extracting text structure features through a double-affine mechanism and a graph neural network; simultaneously using [CLS] word vectors representing semantic information in word-level LSTM and BERT to integrate context information into word vectors of the entity pair, and constructing entity pair features containing context information; and finally, extracting the similarity and difference of sentence features through a similarity-difference relation network to perform small sample relation extraction. According to the method and device, small sample learning is applied in a content big data scene, text features are fully extracted, differences among different types of entity relations are described, and the accuracy of relation extraction is improved.

The invention discloses a knowledge tracking method and system fusing question difficulty. The method comprises the following steps: acquiring a question qt and an answer rt answered by a learner at a t moment and interaction information when the learner answers the question; obtaining a concept matrix Mk and a concept mastering matrix at the moment t, and calculating question difficulty dt according to the interaction information; obtaining the relevancy wt (i) of the question qt and the ith concept; calculating to obtain the overall mastering degree kst of the learner to the concepts related to the question qt; outputting the prediction accuracy prt of the question qt answered by the learner and the prediction difficulty pdt of the question qt for the learner; training a prediction model, wherein the training target is to minimize the difference between the prediction accuracy prt and the answer rt and the difference between the prediction difficulty pdt and the question difficulty dt; and updating the concept mastering matrix of the next time step. While predicting the accuracy of the learner, the invention also predicts the difficulty of the question to the learner, and enhances the expression of the model in the aspect of the knowledge mastering state of the learner.

An article image re-identification method based on multi-feature information capture and correlation analysis uses a convolutional layer with a space attention mechanism and a channel attention mechanism to weight an input feature map, considers the effective combination of channel and space information, not only can pay attention to important features, but also can suppress unnecessary features, and the representation of the concerns can be improved, so that better features are obtained. By using a transformer, the features after image partitioning can be better processed by utilizing a multi-head attention mechanism, richer feature information can be captured, and the correlation among the features can be considered, so that good performance can be obtained, and the efficiency of object image retrieval can be improved. According to the method, the convolutional layer with the channel attention mechanism and the space attention mechanism and the transformer with the multi-head attention mechanism are combined, relatively important features can be paid attention to globally, and fine-grained features can be better captured, so that the re-recognition performance can be well improved.

The invention discloses a text similarity calculation system and method based on a distance perception self-attention mechanism and multi-angle modeling, for a text P and a text Q, text similarity calculation is completed by means of Word2vec, BiLSTM, the distance perception self-attention mechanism and multi-angle modeling, firstly Word2ve is used to obtain word representations of the text P and the text Q, and context representations are captured through the BiLSTM; the method comprises the following steps: firstly, extracting a text P and a text Q, then further extracting text features by utilizing a distance-aware self-attention mechanism, then carrying out interactive modeling on the text P and the text Q by utilizing an interactive attention mechanism, and finally, carrying out similarity modeling on the text P and the text Q from different levels by proposing a multi-angle similarity modeling method so as to obtain a final similarity. According to the method, through a self-attention mechanism based on distance perception and a multi-angle similarity modeling architecture, the text semantic modeling capability is enhanced, the accuracy of text similarity calculation is improved, and the model is simple in structure, efficient in performance, high in robustness and easy to popularize and use in practice.

The invention relates to a saturation and free water surface calculation prediction method based on microscopic heterogeneity characterization. The problem that an existing conventional method cannotaccurately calculate the original saturation of an oil reservoir transition zone is mainly solved. The method mainly comprises the following steps: 1) collecting related data of a target oil reservoir; 2) achieving fitting of a Thomeer function curve and a mercury injection curve to obtain microscopic heterogeneity characterization parameters; 3) comparing and analyzing nuclear magnetic data and the like to confirm microscopic pore throat structure characterization parameters; 4) realizing nonlinear regression through deep research of a coring well mercury injection curve sample point, and obtaining a formula for predicting microscopic pore throat structure characterization parameters by using conventional logging information; 5) inputting related data, and iteratively calculating the water saturation and the free water surface depth; and 6) counting and analyzing a prediction result, and carrying out quality control and confirmation on application of the technology. The method can calculate the reservoir saturation and the free water surface depth accurately, and establishes a bridge understanding from microcosmic to macroscopic and understanding from static to dynamic.

The invention relates to a color image zero-watermarking algorithm based on ternary number decimal order polar complex exponential transformation, which comprises the following steps: a, calculating ternary number decimal order polar complex exponential transformation of an original color image; b, constructing an amplitude sequence through copying and expanding; c, carrying out binarization processing on the amplitude sequence to obtain a binarized amplitude sequence; d, changing the binary amplitude sequence into a binary feature image; e, embedding the LOGO image into the binary feature image by using an XOR operation to obtain a zero-watermark image; and f, finally carrying out zero watermark verification. According to the method, a ternary number theory is combined, ternary number decimal order polar complex exponential transformation is constructed, the relationship among three channels of the color image is utilized, information redundancy is avoided, and the expression and calculation efficiency of the color image is improved. And the constructed ternary number decimal order polar complex exponential transformation is applied to the color image reconstruction and zero watermark algorithm, so that the good stability of the ternary number decimal order polar complex exponential transformation is verified.

The invention discloses an e-commerce comment sentiment analysis model based on part-of-speech features and viewpoint features in combination with a convolutional neural network. The method comprisesthe following steps: step 1, formulating rules by utilizing part-of-speech, dependency grammar analysis and semantic dependency analysis to extract viewpoint features; step 2, introducing part-of-speech features and viewpoint features by adopting a vector splicing method on the basis of word vector representation; and step 3, taking a word vector and an extended feature vector as two input channels of a convolutional neural network to perform sentiment analysis. The problem that comments are inconsistent with scores of the comments is solved, and then merchants are helped to improve the service quality and upgrade the product performance.

The invention discloses a joint pre-stack elastic inversion parameter and deep network target inversion technology, and relates to the technical field of target inversion. The target inversion technology comprises the following specific steps of: obtaining pre-stack elastic impedance inversion parameters; obtaining pre-stack parameters of a division angle; taking well bypass data as input and target data as output constraint; establishing a training network; obtaining a mapping network through training, and directly carrying out mapping calculation on all-region data; and inputting all-region data parameters, and directly obtaining a target data body through the mapping network. According to the method, seismic target parameters can be better predicted, the prediction accuracy and efficiency are improved, data internal information is deeply mined through the deep network, the characterization degree of the information is improved, accumulative errors and approximate errors existing in the process of solving target data through fitting are solved, and the credibility and authenticity of the algorithm are enhanced directly through parameters and different angle data obtained through pre-stack inversion.

The invention provides a deformed face detection method based on multidirectional fusion attention for face deformation detection, and the method comprises the following steps: 1) segmenting and normalizing the face of an image according to eye coordinates detected by a dlib mark point detector; 2) considering the position information ignored by channel attention, and proposing a new attention module; and 3) a double-branch convolutional network is fused, so that the detection precision is improved. And (4) classifying the final feature map by using an SVM (Support Vector Machine).

The invention relates to an inter-color correlation-based stereo matching sparse block aggregation strategy method. The method is characterized by comprising the following steps of: obtaining a groupof test images from a Middlebury data set platform and transmitting the test images into a computer 1; improving a matching cost function by utilizing correlation between a pixel and a color space soas to better express local properties of the pixel; constructing a mean value block-based sparse aggregation strategy to improve the algorithm efficiency; and eliminating mis-matched points by utilizing left-right consistency detection and median filter. Through the method, the matching precision in weak texture areas is improved and the operation time is shortened.