39results about How to "Differentiated" patented technology

The invention discloses a multilevel semantic feature-based face feature extraction method and recognition method. The method includes the following steps that: 1) organ areas of each image in a facial image set A are divided; 2) bottom-level features of each organ are extracted and clustered; two clusters are extracted from clustering results and are adopted as positive and negative samples, and the positive and negative samples are trained in a paired combination manner such that a classifier set can be obtained, and the results of discrimination which is performed on the bottom-level features by the classifier set are united so as to obtain the middle-level features of the organ; the images in the A are the classified according to tags; any two classifications are selected from classification results of the tags and are adopted as positive and negative samples, and the positive and negative samples are trained in a paired combination manner such that a classifier set can be obtained, and the results of classification and discrimination which are performed on all the middle-level features in the A by the classifier set are united so as to obtain high-level features of the tags; the bottom-level features, the middle-level features and the high-level features are adopted to construct face features of the images; face features Vq are generated for any image q to be searched; and the face features Vq are matched with the face features in the A, and query results are returned. With the multilevel semantic feature-based face feature recognition method and recognition method adopted, recognition accuracy and stability can be improved.

The invention provides a pedestrian re-identification method based on a multi-component self-attention mechanism. The method comprises the following steps: firstly, pre-training a deep convolutional neural network backbone model; then, after the backbone model is branched, a multi-component self-attention network is constructed, and multi-component self-attention characteristics are obtained; inputting the multi-component self-attention features into a classifier, and performing joint training to minimize cross entropy loss and metric loss; and finally, inputting a test set picture into the trained model, fusing the output part features to obtain an overall feature, and realizing pedestrian re-identification through metric sorting. According to the method, various challenges existing in the pedestrian re-identification problem are fully considered, a multi-component self-attention mechanism is provided, the attention activation area is effectively expanded, and pedestrian features areenriched; the self-attention module enables the network to pay more sufficient and careful attention to an area with discrimination characteristics, and the spatial attention module and the channel attention module are integrated into the network in a residual form, so that the network is more robust and stable, and is easy to train.

The invention discloses a speech recognition decoding method integrating guide probability and provides a guide probability model for overcoming the insufficiency that the traditional speech recognition system insufficiently utilizes position information of a speech frame in an acoustic characteristic space, and for the purpose of describing the probability that the speech frame belongs to different parts of the acoustic characteristic space and guiding a decoding process. The method comprises the following steps of training a universal background model for describing the whole acoustic characteristic space, computing a main gaussian component of the speech frame in the universal background model, utilizing an acoustic model of a recognition system to forcibly segment a training corpus, obtaining a phoneme of the speech frame, counting response frequency of the phoneme and main gaussian, normalizing the respond frequency, obtaining the guide probability, fusing the guide probability to total score computation of a speech recognition path, and guiding a decoder to enhance or weaken the path.

The invention discloses a handwriting recognition method and a handwriting recognition system based on WIFI channel state information. The method comprises the following steps of S1, using a data acquisition module to collect channel state information, processing the collected channel state information by using a movement section detection and segmentation method, and using a K-nearest neighbor algorithm and a dynamic time normalization method to train a classifier, thus acquiring a trained classifier; and S2, using the data acquisition module to acquire the collected channel state information, recognizing the collected channel state information by using the trained classifier, and thus acquiring a recognition result. According to the method and system provided by the invention, the handwriting letter recognition under the WIFI environment is achieved by using the channel state information of the wireless signal, the limitation that traditional behavior recognition needs the user to additionally carry special equipment is overcome, only the existing common consumption level WIFI equipment is needed, and the additional equipment overheads are reduced.

The invention belongs to the technical field of cross-modal data retrieval, and particularly relates to a cross-modal hash retrieval algorithm based on a fine-grained similarity matrix. The algorithmprovided by the invention mainly aims at two tasks of image retrieval texts and text retrieval images, and comprises the following steps: hash code reasoning: constructing a fine-grained similarity matrix by utilizing label information of image text pairs, so that hash codes reserve fine-grained similarity information between image text data items; constructing an auto-encoder to enable the hash code to reserve semantic information in the label as much as possible; hash function learning: training two Hash functions, mapping images and texts to Hash codes respectively, wherein target functionsused by Hash code learning include Hash code mapping loss, similarity retention loss with weight and classification loss. The invention has relatively high retrieval precision in two tasks of image search texts and text search images.

The invention relates to a sun screening agent hydrogel composition, and a preparation method and an application thereof. The composition comprises 5-15wt% of chitosan, 5-20wt% of glycerophosphate, 2-10wt% of sodium hyaluronate, 5-15wt% of a sun screening agent and 50-70wt% of water. Compared with compositions in the prior art, the composition disclosed in the invention has the advantages of stable cladding of the sun screening agent, prevention of direct contact with human body skins, effective avoiding of irritation and allergy effects on the human body skins, and substantial improvement of the ultraviolet protection, anti-oxidation and whitening effects of sun screening products.

The invention discloses a vehicle re-identification method and system using weak supervision area recommendation, and the method comprises the steps: collecting vehicle image data captured by a road monitoring camera, carrying out the data enhancement of the vehicle image data, and applying the data to a weak supervision area recommendation network designed in the invention; designing the structure of a weak supervision region recommendation network model, and optimizing a region recommendation module by using classification loss and region suggestion loss functions, so that a multi-scale local region with rich information degree in an image can be extracted, and local region features with higher discrimination and description are obtained, and the local features and the global features are optimized by using triple loss, so that the identification capability of the network is enhanced; and using an adaptive gradient descent algorithm to supervise the training of the weakly supervised region recommendation network in an end-to-end manner, after a trained model is obtained, calculating the similarity between the vehicle query set and the image library set, and sorting according to the similarity to obtain a vehicle re-identification result.

The invention discloses an image detection method and device. The image detection method comprises the steps: obtaining a to-be-detected image, extracting a model according to a preset multi-scale feature, extracting multi-scale feature information of the to-be-detected image, carrying out dimension reduction on the multi-scale feature information; obtaining target image feature information; obtaining target image feature information according to the target image feature information, searching historical image feature information according to the target image feature information to obtain a search result, if the historical image feature information matched with the target image feature information exists in the search result, calling a historical detection record corresponding to the historical image feature information, and outputting the historical detection record as a detection result of the to-be-detected image. According to the image detection method, repeated detection is avoided, and richer features can be extracted, and the image can be better described.

The invention belongs to the field of computer vision and pattern recognition, relates to a fine classification method and system based on graph network multi-granularity feature learning and aims tosolve the problem of low image fine classification accuracy due to insufficient utilization of a target multi-granularity level relationship in the prior art. The method comprises the steps of: constructing a multi-granularity level relationship graph based on an obtained to-be-classified image; extracting the feature of each granularity level in the multi-granularity level relationship graph through a multi-granularity level feature extraction network; carrying out level relationship embedding on each granularity level feature in a multi-granularity level feature set; and obtaining a prediction category of each granularity level through a classifier. According to the method and system, a graph neural network is used for learning the semantic relation between the multi-granularity levels of an image, so that feature learning of all granularities is improved in a common promotion mode; the feature extraction network is composed of a main network and three branch networks, so that a model parameter quantity is reduced while high precision is achieved.

The invention provides a label-based document classification method, system and device and a storage medium; the method comprises the following steps: obtaining a to-be-classified text and a label, and carrying out the word segmentation of the to-be-classified text to obtain a word embedding vector; performing word segmentation on the text content of the label to obtain a label embedding vector; combining the word embedding vectors, and determining a first embedding sequence; obtaining a statement embedding vector according to the label embedding vector and a word weight coefficient in the first embedding sequence; combining the statement embedding vectors, determining a second embedding vector, and obtaining a text embedding vector according to the tag embedding vector and a statement weight coefficient in the second embedding sequence; determining the classification probability of the to-be-classified text according to the text embedding vector, and performing classification according to the classification probability. According to the method, the document representation vector adopted by the method is more distinctive, so that more excellent classification performance can be obtained; the method can be widely applied to the technical field of natural language processing.

The present invention relates to a millimicrosphere composite medicine containing mixture of alprostadil and vitamin F and its preparation method. The invented medicine can greatly prolong the half-life time in vivo and has good target action for tumor.

The invention discloses a lithology prediction method based on the residual network. Lithology qualitative analysis is achieved, the internal relation between the seismic attribute and the lithology is sought by utilizing the characteristic that the residual network can build a deeper convolutional network, after a prediction model is trained, corresponding lithology categories can be obtained according to a seismic attribute forming characteristic pattern, the quantity of output labels, that is, the quantity of output layer neurons is determined according to the quantity of the lithology categories in data, the output value of each neuron represents the probability that the group of data belongs to the corresponding lithology category, and the lithology characteristics of the neurons canbe expressed more precisely.

The invention provides a black phosphorus composition, a biological material containing the black phosphorus composition, a preparation method and application. The black phosphorus composition comprises the components, in percentages by weight, of 69.901% to 89.999% of a biocompatible material, 0.001% to 0.099% of black phosphorus and 10% to 30% of a bioadhesive substance. The biological materialcontaining the black phosphorus composition comprises a catheter, wherein the catheter is prepared from the black phosphorus composition; and a bioadhesive substance layer is arranged on the side wallof the catheter. The preparation method of the biological material comprises the steps of carrying out 3D printing, mold method or electrostatic spinning and the like on the black phosphorus composition to prepare the catheter, spraying a layer of bioadhesive substance inside the catheter by adopting 3D printing or electrostatic spinning, and carrying out crosslinking curing to obtain the biological material. The prepared catheter can promote regeneration, proliferation, differentiation and the like of nerves, muscles, bones, blood vessels and the like. In-vivo experiments show that the repair of animal peripheral nerve, muscle and the like, such as sciatic nerve and muscle, is superior to autologous nerve transplantation, and has practical and clinical application values.

The invention discloses a bearing fault monitoring method and device. The method comprises the steps of obtaining and training full life cycle historical data of a to-be-monitored bearing; processing the first alarm time data and the first early warning time data to obtain first alarm time data and first early warning time data; different preset large cycles and different preset small cycles are respectively selected, and are gradually converted into a plurality of specific value different cycles, and second alarm time data and second early warning time data are determined; substituting the first alarm time data, the first early warning time data, the second alarm time data and the second early warning time data into a preset different-period evaluation function to obtain optimal data; and if the false alarm exists, updating and monitoring the bearing to be monitored by using the optimal alarm data and the optimal early warning data. According to the bearing fault monitoring method and device provided by the embodiment of the invention, the alarm data and the early warning data are dynamically updated by designing the specific bearing fault monitoring strategy, so that the timeliness and the accuracy of bearing fault monitoring of mechanical equipment are improved.

The invention discloses a bone cement composite material and a preparation method thereof, and relates to the technical field of medical materials. It includes: solid phase components and liquid phase components, the mass ratio of the solid phase components and liquid phase components is 2:1; wherein, the solid phase components include: methyl methacrylate copolymer, paraffin / Polymethyl methacrylate-methylenebisacrylamide phase change microcapsules, in the solid phase component, the mass fraction of methyl methacrylate copolymer is 70-90%, paraffin wax / polymethyl methacrylate The mass fraction of the ester-methylenebisacrylamide phase-change microcapsules is 10-30%; the liquid phase component includes methyl methacrylate. The raw material of the bone cement composite material of the present invention has high biological safety, the paraffin wax / polymethyl methacrylate-methylenebisacrylamide phase change microcapsules have high stability themselves, and have good combination with acrylic resin bone cement, which can effectively reduce the Tissue heat damage increases the safety of acrylic polymer bone cement used in humans.

The invention provides a black phosphorus composition, a biological material containing the black phosphorus composition, a preparation method and application. The black phosphorus composition comprises the components, in percentages by weight, of 69.901% to 89.999% of a biocompatible material, 0.001% to 0.099% of black phosphorus and 10% to 30% of a bioadhesive substance. The biological materialcontaining the black phosphorus composition comprises a catheter, wherein the catheter is prepared from the black phosphorus composition; and a bioadhesive substance layer is arranged on the side wallof the catheter. The preparation method of the biological material comprises the steps of carrying out 3D printing, mold method or electrostatic spinning and the like on the black phosphorus composition to prepare the catheter, spraying a layer of bioadhesive substance inside the catheter by adopting 3D printing or electrostatic spinning, and carrying out crosslinking curing to obtain the biological material. The prepared catheter can promote regeneration, proliferation, differentiation and the like of nerves, muscles, bones, blood vessels and the like. In-vivo experiments show that the repair of animal peripheral nerve, muscle and the like, such as sciatic nerve and muscle, is superior to autologous nerve transplantation, and has practical and clinical application values.

The invention provides an image visual attribute mining method based on sparse factor analysis. The technical solution includes the following steps: the first step is to calculate the feature matrix of the image collection. Calculate a D-dimensional gradient orientation histogram feature vector for each image Ii in the image set I, and obtain the feature matrix X of the image set. In the second step, the visual properties of the image collection are mined. Initialize the visual attribute matrix A as the first K columns of X, initialize the visual attribute mixing coefficient matrix Y as a unit diagonal matrix, and use alternate optimization to iterate to obtain the optimal visual attribute matrix A. The invention automatically mines image visual attributes, avoids the disadvantage that a large number of manually marked samples are required for attribute learning, and greatly improves the distinguishing performance of visual attributes by introducing sparse constraints when optimizing the visual attribute mixing coefficient matrix.

The invention discloses a gravure roller coating and knife coating alternate printing technique, including: subjecting fabric to gravure roller printing prior to pre-drying, flattening with a presserroller under the temperature of 100-130 DEG C and the pressure of 2-2.5 MPa, drying, performing knife coating, and drying to obtain a finished product. The manufactured fabric has the differential andspecial printing features of dryness, coolness, softness and smoothness, has excellent color fastness, has effectively improved grade and added value and has improved competitiveness.

The invention provides an image visual attribute excavation method based on sparse factor analysis. The technical scheme comprises following steps of firstly calculating a characteristic matrix of an image set: calculating a histogram characteristic vector in a D-dimensional gradient direction of each of images Ii in the image set I so as to obtain the characteristic matrix of the image set; and secondly: excavating visual attributes of the image set: initializing a visual attribute matrix A into the front K line of X, initialing a visual attribute mixing coefficient matrix Y into a unit diagonal matrix, and carrying out iteration via alterative optimization so as to obtain the optimal visual attribute matrix A. According to the invention, image visual attributes are automatically excavated, so a disadvantage of demanding a large number of manual marking samples of attribute learning is overcome; and by intruding sparse constraints during optimization of the visual attribute mixing coefficient matrix, distinguishing performance of the visual attributes is greatly improved.

The present invention provides a class-level 6D pose and size estimation method and device, which perform 6D pose and size estimation based on the depth observation data of the target object in the image and the internal parameters of the camera that captures the image, and is characterized by comprising the following steps: Based on the camera internal parameters, the depth observation data is reprojected to three-dimensional space and normalized to obtain part of the observation point cloud data of the target object; according to the target object, the normalized template point cloud data of the same type as the target type is selected from the template point cloud database; Part of the observed point cloud data and the normalized template point cloud data use the geometric feature learning network to estimate the 6D pose and size of the target object, and obtain the corresponding predicted pose and predicted size. The class-level 6D pose and size estimation method and device of the present invention can accurately estimate the 6D pose and size of instance objects whose classes are known but not seen, and have high accuracy, generalization ability and practicability.

The invention discloses a CFCC space gradient-based keyboard single-key keystroke content identification method and belongs to the technical field of non-contact identification. The method is dividedinto an initial identification stage and a gradient identification stage. In the initial identification stage, a CFCC value of a keystroke sound signal is extracted. Keystroke content is recognized through a BP neural network method, the Manhattan distance and CFCC difference between keys with low recognition accuracy and other keys are calculated through a gradient recognition stage, a CSG matrixis constructed, a new training set and a new test set are constructed, the BP neural network is retrained, and a final classification result is obtained. The method can better alleviate the interference of the environment and the influence caused by the diversity of measurement equipment, is more robust than an original CFCC, enables the features of a sound signal to be differentiated in space, and is more stable in time.

The invention discloses an application method in zero-order learning based on a correlation double autoencoder. The present invention establishes autoencoders for visual features and semantic features respectively. But these two autoencoders are not independent, they are connected. We add the encoded visual features to the encoded semantic features, and then decode the added semantic features. Finally, the decoded semantic features are added to the original semantic features to obtain better and more complete semantic features. The optimized semantic features are then mapped to visual features for classification and recognition. The present invention optimizes the semantic features by using the correlation double autoencoder model to obtain more distinguishable and finer-grained semantic features. The optimized semantic features obtained in this way are then mapped to the visual feature space, which can achieve better classification and recognition accuracy.

The invention discloses a probability linear speaker-distinguishing identifying method based on priori knowledge structured covariance, which is capable of structuring a covariance hypothesis and iterative process of a probability linear identifying-analyzing model based on random useful information related to training voice; and finally, a probability linear identifying-analyzing model that can be more distinctive and can reflect the real situation can be trained; and at the same time, two structuring coefficients are introduced to make the model adjustable and can be self-adaptive to be optimum aiming to various different structuring information. By adopting the model trained by this method provided herein, compared with the traditional model, the evaluating effect of identifying the speaker on the same dataset is improved significantly; and the equal error rate (EER) and the minimum detect error cost function (norm minDCF) can be lowered 10 percent to 20 percent in the evaluating database of identifying an internationally authoritative speaker.

The invention relates to a cross-modal retrieval method and device, a network training method and device, equipment and a medium. The cross-modal retrieval method comprises the steps of obtaining to-be-retrieved data and candidate data; the data to be retrieved and the candidate data correspond to different modes; extracting a first feature of the to-be-retrieved data and a second feature of the candidate data based on a cross-modal retrieval network; and retrieving data matched with the to-be-retrieved data from the candidate data according to the matching degree of the first feature and the second feature. By adopting the cross-modal retrieval network disclosed by the invention, the input local information of the to-be-retrieved data and the candidate data can be accurately captured, more effective features are output, and the more effective features are more distinctive in the same mode and are more identifiable in different modes, so that the fine-grained cross-modal retrieval performance is improved.

The invention discloses a fentanyl substance detection method and device based on a twin network. And acquiring mass spectrum data of a to-be-detected substance, carrying out standardization treatment, and classifying the to-be-detected substance by using the fentanyl substance detection model. When the fentanyl substance detection model is trained, the detection model is formed by a twin network and a classification network, and the twin network comprises two feature extraction networks sharing weight; during testing, any feature extraction network in the trained twin network is deleted and then cascaded with the classification network. And the sample pair is used as the input of the twin network, so that the sample data for model training is greatly increased, and the small sample problem of a supervised learning method in a fentanyl substance detection task is effectively solved. In the design of the loss function, a comparison loss function, maximization of inter-class difference and minimization of intra-class difference are added, so that the extracted features are more distinctive, and the performance of the detection model is improved.

The invention relates to the technical field of image description, in particular to a reward in an image description model and a construction method of the image description model, and the construction method of the reward comprises the steps: obtaining label description of a test image and generation description of a preset image description model based on the test image; based on the label description and the generation description, calculating a reward of the n-tuple in the generation description; comparing the size relationship between the reward of the n-tuple and the first threshold value with the size relationship between the reward of each word in the n-tuple and the second threshold value so as to screen out the word with the reward to be increased in the generated description; and according to the label description and the reward of the n-tuple corresponding to the word with the reward to be increased, calculating the reward of the word with the reward to be increased. According to the method, higher rewards are given to words which are not frequently appeared to realize constraint, so that the problem of non-uniform word distribution can be well solved, more specific visual details of an image can be captured, and the description accuracy of the image can be improved.

The invention discloses an abnormal detection method for a gas circuit of an aero-engine based on deep learning and Gaussian distribution and relates to an abnormal detection method for a gas circuit of an aero-engine. The invention aims to solve the problems that in an existing abnormal detection method for the gas circuit of the aero-engine, QAR data are not widely applied, the false alarm rate of abnormal detection of the engine is high and the accuracy is low. The method comprises the following steps: I, selecting a parameter set in the QAR data, wherein the data set comprises a performance parameter of the gas circuit of the engine and an external environmental parameter; II, calculating the difference value of the performance parameters of two engines on a same plane in the parameter set, and forming a novel parameter set by the difference value and the external environmental parameter; III, extracting data characteristics of the novel parameter set in the step II by using an accumulating and noise-eliminating automatic coder model in a deep learning method; and IV, performing abnormal detection on the data characteristics obtained in the step III by a density estimation algorithm based on Gaussian distribution to obtain a result. The method disclosed by the invention is used in the technical field of fault diagnosis of the aero-engine.

The invention discloses a production method of an antibacterial quick-drying super cotton-like knitted fabric, and relates to the technical field of antibacterial quick-drying fabrics, and the production method comprises the following steps: step 1, preparing materials; 2, the materials are distributed according to the proportion; step 3, putting into a machine for processing; a1, cloth fibers are staggered, overlapped and knitted; a2, cleaning the cloth; a3, drying the cloth; a4, collecting the cloth; and step 5, filling cloth for sale. An entering water source is differentiated through the two-way differentiation plate, the water source is filtered in cooperation with the double filtering plate, the differentiated water source is atomized and sprayed out through the multi-pipe spraying head on the surface of the connecting plate, the water source has the advantages of being differentiated, filtered and atomized, and the problem that if the water source is not atomized, the water source cannot be atomized is solved. The problem that the cloth is blown away from the original position by a high-speed water source is solved, and the effects of differentiating, filtering, atomizing and spraying the water source are achieved.

The invention belongs to the technical field of cross-modal data retrieval, in particular to a cross-modal hash retrieval algorithm based on a fine-grained similarity matrix. The algorithm of the invention is mainly aimed at two tasks of image retrieval text and text retrieval image, including: hash code reasoning: using the label information of the image-text pair to construct a fine-grained similarity matrix, so that the hash code preserves the difference between the image-text data items fine-grained similarity information; construct an autoencoder so that the hash code retains the semantic information in the label as much as possible; hash function learning: train two hash functions to map images and text to hash codes respectively, ha The objective functions used in hash code learning include hash code mapping loss, weighted similarity retention loss, and classification loss. The present invention has relatively high retrieval accuracy in both tasks of image search text and text search image.