132results about How to "Strong representation ability" patented technology

The invention discloses a bimodal emotion recognition method and system based on 3D convolutional neural networks. According to the method, first, two 3D convolutional neural networks for face emotionrecognition and body gesture emotion recognition are constructed respectively, and network model parameters are optimized based on a training set and a verification set of a bimodal face and body gesture database; second, the two neural networks obtained after optimization are tested based on a test set of the bimodal face and body gesture database to obtain a face emotion recognition confusion matrix and a body gesture emotion recognition confusion matrix; and last, priori knowledge of the face emotion recognition confusion matrix and the body gesture emotion recognition confusion matrix isutilized to fuse bimodal recognition results of a newly input face video sequence and body gesture video sequence, and a bimodal emotion classification result is obtained. Through the method, the 3D convolutional neural networks and a bimodal fusion algorithm are adopted, the subjectivity of manual feature design is avoided, the limitation of single-modal emotion recognition is overcome, and the accuracy and robustness of emotion recognition can be effectively improved.

The invention relates to a Fourier descriptor and gait energy image fusion feature-based gait identification method. The method comprises the steps of performing graying preprocessing on a single frame of image, updating a background in real time by using a Gaussian mixture model, and obtaining a foreground through a background subtraction method; performing binarization and morphological processing on each frame, obtaining a minimum enclosing rectangle of a moving human body, performing normalization to a same height, and obtaining a gait cycle and key 5 frames according to cyclic variation of a height-width ratio of the minimum enclosing rectangle; extracting low-frequency parts of Fourier descriptors of the key 5 frames to serve as features I; centralizing all frames in the cycle to obtain a gait energy image, and performing dimension reduction through principal component analysis to serve as features II; and fusing the features I and II and performing identification by adopting a support vector machine. According to the method, the judgment whether a current human behavior is abnormal or not can be realized; the background is accurately modeled by using the Gaussian mixture model, and relatively good real-time property is achieved; and the used fused feature has strong representability and robustness, so that the abnormal gait identification rate can be effectively increased.

The invention discloses a newborn pain degree assessment method and system based on facial expression recognition, and the method comprises the steps: building a newborn pain facial expression image data set which comprises a preprocessed newborn facial expression image and a corresponding expression class label; constructing a DCNN (deep convolutional neural network) for estimating the pain degree of a newborn, employing a disclosed large-size data set with a label for the pre-training of a network, obtaining an initial weight parameter value, carrying out the fine tuning of the network through the facial expression image data set, and obtaining a trained network model; inputting a to-be-tested newborn facial image into the trained network for facial expression classification and recognition, and obtaining a pain degree assessment result. The method can make the most of the features extracted through the DCNN, can obtain a better pain degree assessment result through a small-size newborn pain degree facial expression image data set, and is a new method for the development of a system for automatic assessment of the pain degree of the newborn based on the facial expression recognition.

The present invention discloses a polarization SAR ground object classification method based on a self-step learning convolutional neural network, in order to mainly solve the problems that the priorart has low accuracy in classifying complex ground object scenes and is heavily affected by noise. The implementation scheme comprises: 1, obtaining a polarization scattering matrix S and a pseudo color RGB image under the Pauli basis from original complete polarization SAR data; 2, constructing a three-dimensional matrix to form a sample set for each pixel, and constructing a training sample setand a test sample set; 3, constructing a convolutional neural network and training the convolutional neural network based on self-step learning to accelerate network convergence and improve the generalization ability of the network; and 4, classifying the test samples by using the trained convolutional neural network to obtain a final complete polarization SAR ground object classification result.According to the method disclosed by the present invention, accuracy for classifying the target ground objects of complex ground object scenes in the polarization SAR image is improved, and the methodcan be used for feature classification and target recognition.

The invention discloses a pyramid network Chinese herbal medicine identification method based on an attention mechanism, which comprises the following steps: 1) constructing a Chinese herbal medicinedata set, and making a Chinese herbal medicine training set and a Chinese herbal medicine test set; 2) constructing a feature fusion structure block based on a channel attention mechanism, and introducing a competition attention module; 3) adding a spatial attention mechanism to the feature fusion structure block of the pyramid network, adjusting the two information flows by using a spatial collaborative attention module, and fusing the two adjusted information flows as output; 4) constructing a pyramid network based on an attention mechanism, and training by using a Chinese herbal medicine training set; and 5) transmitting the pictures in the Chinese herbal medicine test set to the trained network model to identify the Chinese herbal medicine types corresponding to the pictures, thereby improving the Chinese herbal medicine identification accuracy and performance, assisting related industrial personnel to identify the Chinese herbal medicines, and facilitating non-professionals to identify the Chinese herbal medicines.

The invention discloses a pedestrian detection method based on multi-feature fusion. The method includes the following steps that noise suppression is conducted on an obtained depth image; area-of-interest detection is achieved through a depth threshold; on the basis of obtaining the HOG-LBP union features, target detection is achieved through a classifier. According to the pedestrian detection method based on multi-feature fusion, the fused features of the HOG and the LBP are classified through a support vector machine. The method has higher pedestrian representation capacity, and the pedestrian detection accuracy is obvious improved under a complex background.

The present invention discloses a newborn pain expression recognition method and system based on a deep 3D residual network. The method comprises: establishing a newborn expression video library containing pain expression category tags, and dividing samples in the newborn expression video library into a training set and a verification set; constructing a deep 3D residual network for newborn pain expression recognition, pre-training the network by using a publicized large-scale video database with category tags to obtain initial weight parameter values, and then performing fine-tune on the network by using the samples in the training set and in the verification set in the newborn expression video library to obtain a trained network model; and inputting a to-be-tested newborn expression video segment into the trained network model for expression classification recognition, and obtaining a pain expression recognition result. According to the technical scheme of the present invention, a deep 3D residual network is used to extract temporal and spatial dynamic features capable of reflecting time information from the video, and the change of the facial expression can be better characterized, so that the accuracy of the classification recognition can be improved.

The invention discloses a hperspectral remote sensing image classifying and recognizing method which is used for solving the technical problem that an existing hperspectral remote sensing image classifying method based on sparse representation and spectrum information is low in precision. According to the technical scheme, by means of combination of a word bag model and hperspectral remote sensing image data, various professional dictionaries are generated, a sparse representation dictionary is further constructed, then sparse representation characteristics of image elements are calculated according to the dictionary, sparse representation coefficients of the image elements are restricted in a space dimension by means of space continuity, and finally hperspectral image classification is achieved by solving an optimization problem. Representational capacity of the generated dictionary is relatively strong, and space structure information of hperspectral images is fully considered, so that classification precision is improved. The AVIRIS hperspectral images are classified, the overall classification precision is improved from 82.58% in the prior art to 86.87%, processing time is shortened from 97.469 seconds in the prior art to 35.539 seconds, and efficiency is improved by nearly three times.

The invention relates to the technical field of image processing, and aims to provide a pedestrian flow statistics method based on deep learning and multi-target tracking. The method mainly comprisesthe following steps: S1, shooting a pedestrian monitoring video and reading an image in the video; S2, setting an effective area and a flow count of the image; S3, constructing a pedestrian detectionmodel based on deep learning and training the pedestrian detection model; S4, performing current pedestrian detection to obtain coordinates and image blocks of a current pedestrian frame; S5, trackingthe current pedestrian by using a multi-target tracking algorithm based on deep learning, and generating coordinates of the current pedestrian; S6, generating a moving track of the current pedestrian; S7, judging whether the current pedestrian leaves the effective area or not; If yes, entering the step S8, and if not, entering the step S4; S8, a noise threshold value is selected, and noise judgment is carried out; And S9, deleting the coordinates of the current pedestrian in the continuous video frames. According to the invention, an accurate flow statistics result can be provided in an actual use scene.

The invention discloses a polarization-texture characteristic and DPL (Dictionary Pair Learning)-based polarimetric SAR (Synthetic Aperture Radar) image classification method, and mainly solves the problems of long calculation time and low classification efficiency caused by non-comprehensive image information and poor dictionary distinguishing performance in a polarimetric SAR image classification process in the prior art. The method comprises the following specific steps of (1) reading polarimetric SAR images; (2) performing filtering; (3) creating a sample set; (4) performing sparse coding; (5) creating a neighborhood characteristic sample matrix; (6) normalizing the neighborhood characteristic sample matrix; (7) selecting a training sample and a test sample; (8) training a synthetic dictionary and an analytic dictionary; (9) testing the synthetic dictionary and the analytic dictionary; (10) performing coloring; and (11) outputting a classification result graph. The method has the advantages of high classification correctness and high classification efficiency of the polarimetric SAR images.

The invention discloses a dynamic expression recognition method and system based on an attention mechanism between space-time streams. The method comprises the steps of firstly collecting facial expression video clips, and establishing a facial expression video library containing expression category labels; then constructing a double-flow convolutional neural network model embedded with an attention mechanism module between space-time flows, wherein the model comprises a data processing layer, a space flow branch, a time flow branch, the attention mechanism module between the space-time flows, a feature fusion layer, a full connection layer and a classification layer; training the model by using video samples in a facial expression video library; and finally, performing facial expression recognition on a newly input video by using the trained model. According to the method, the attention mechanism module between the space-time streams is embedded in the double-stream convolutional neural network, so that the information interaction between the space domain features and the time domain features can be realized, the dynamic association information between the space domain features and the time domain features is captured, the features with high discrimination capability are obtained, and the accuracy and robustness of facial expression recognition are improved.

A non-rigid target detection method comprises the following steps of performing like-object detection on an input image and finding out rectangular areas which are possibly provided with like-object targets in the input image; performing characteristic extraction on every obtained rectangular area through a pre-trained non-rigid target model and obtaining HSC characteristics and textural features of every rectangular area; forming into a characteristic pyramid; performing PCA projection on the characteristic pyramid and classifying modes on characteristic pyramid space at the low latitude; convolving every scale of the pyramid through every root and obtaining a root scoring chart of every layer of the pyramid; convolving every corresponding part layer of the root layer through part filters and obtaining a root score chart of all parts; performing distance transformation to reconstruct all part scores of every root, comparing and judging a threshold value of the same model and obtaining a position of a non-rigid target of the rectangular areas. The non-rigid target detection method is good in detection effect, rapid in speed and low in leak detection rate.

The invention discloses a multi-modal fusion sign language recognition system and method based on graph convolution. The system comprises a feature extraction module, a feature fusion module, a sequence learning module and an alignment translation module. The method comprises the following steps: 1, respectively extracting color, depth and skeleton features of video frames from a sign language video database by using a convolutional neural network and a graph neural network; 2, combining the multi-modal features, and fusing the network fusion features through a multi-modal sequence; 3, constructing a bidirectional recurrent neural network to perform sequence learning on a series of fused fragment-level features; and 4, aligning the feature sequence through the connected subjective timing sequence classification model, and translating a complete sign language sentence. According to the invention, translation of continuous sign language sentences can be realized, and the accuracy of continuous sign language translation is improved.

The invention provides a human face detection method and device, which belongs to the technical field of image recognition. The method comprises the steps that the position response feature map of a human face correction candidate area in an original image is acquired; the position response feature map is divided to acquire the corresponding number of squares, and the square type of all squares is distinguished according to the position of each square; the weight corresponding to each square type is determined according to the human face detection contribution degree corresponding to each square type; the area feature vector corresponding to the human face correction candidate area is calculated based on the feature vector and weight corresponding to each square; and based on the area feature vector, an area containing a human face in the original image is output. The influence of external conditions on detection results in complex scenes can be reduced by introducing the weights of different human face parts for different square types in the position response feature map. The accuracy of human face detection is high, and the applicability of human face detection is improved.

The invention provides a human detection method based on non-negative matrix factorization, which mainly solves the problems of high feature dimension and high computational load of the existing method. The human detection method is implemented by the following processes of: extracting a scale and rotational invariance SIFT feature point of a training sample image in a CVC-02 database; executing non-negative matrix factorization on the scale and rotational invariance SIFT feature point to get one basis matrix and a coefficient matrix; orthogonalizing and transposing the basis matrix, then, multiplying the basis matrix with the scale and rotational invariance SIFT feature point of the training sample image to get the feature of the training sample image; executing classification training on the obtained feature by an SVM (Support Vector Machine) classifier to obtain a detection classifier; inputting the features of the to-be-detected image into the classifier, according to the classification result, combining all the scanning windows divided for the human to form the final human detection result. The human detection method has the advantages of low feature dimension and high detection accuracy, and can be used for classification and detection on human bodies and other targets in the image.

The invention discloses an intelligent quay crane key mechanism health state evaluation method. The method comprises the steps that firstly, vibration signals of a quay crane mechanism are preprocessed, multiple time domain features are extracted, feature selection is conducted, and a grid capable of measuring vibration feature data distribution is generated through an unsupervised clustering algorithm; then, a hidden Markov model is adopted to train the generated data distribution grid, mapping between data grid distribution and the mechanism health state is established, and a quay crane state evaluation model is obtained; and finally, inputting the to-be-tested data distribution grid into the state evaluation model, and evaluating the health state of the quay crane mechanism according tothe log-likelihood probability output by the model. According to the method, massive unlabeled data generated in quay crane operation can be effectively processed, the problem of lack of quay crane fault data in practical application can be solved, intelligent evaluation of the health state of a quay crane key mechanism is realized, intelligent maintenance of the quay crane is promoted, and the safety and reliability of quay crane operation are improved.

The invention discloses an affinity prediction and analysis method for a human amphiphysin-1 Src homology-3 domain binding peptide. The method comprises the following steps of: selecting an electrical property parameter, a three-dimensional property parameter, a hydrophobicity parameter and hydrogen bond property parameter of an amino acid, establishing physicochemical property classification scores of the amino acid by adopting a principal component analysis method, representing each amino acid residue in a peptide sequence by using the scores to realize the parameter-based description of a peptide sequence structure, selecting characteristic parameters closely related to affinity from parameters of the peptide sequence structure by using a genetic algorithm, establishing a relational model, namely an affinity prediction and analysis model between the characteristic parameters and the affinity by using a Gaussian process, obtaining structural factors influencing the affinity of the peptide by analyzing hyper-parameters of the model, and performing prediction by using the model to obtain an affinity value of the peptide. By the method, a powerful tool can be provided for the high-throughput screening of high-affinity human amphiphysin-1 Src homology-3 domain binding peptides, and development and researches on human amphiphysin-1 Src homology-3 domain binding peptide inhibitors are further prompted.

The invention provides a short video recommendation method, which comprises the following steps of: constructing a multi-source heterogeneous graph, and extracting data features from data of different modalities; constructing a hierarchical graph neural network model, and inputting the data features into the hierarchical graph neural network model for training; and using the trained hierarchical graph neural network model to recommend a short video to a target user. By combining data features extracted in different modes, a social network is comprehensively considered to construct a multi-source heterogeneous graph; the hierarchical graph neural network model constructed on the basis of the heterogeneous graph can capture different connection relationships in the user, the short video and the tag and association relationships among the connection relationships, the customized graph neural network can learn high-level feature representation of the short video and the user, the representation capability is higher, and short video recommendation is facilitated.

The invention discloses a method for calculating and identifying protein kinase phosphorylation specific sites, which is characterized in that, the method comprises the following steps: a) establishing a new protein sequence structure characterization method-amino acid three-dimensional character scoring based on the active ingredient analysis method; b) using the amino acid three-dimensional character scoring for characterizing the structure features of the protein kinase phosphorylation specific sites; c) using a Fisher criterion scoring method for selecting parameters which are closely related to the features of the protein kinase phosphorylation specific sites; and d) and establishing a protein kinase phosphorylation specific site identification model by a radial basis kernel support vector machine, carrying out the self-replacement verification respectively, and using the retaining 1/10 cross-verification and the external verification for verifying the predictive capability of the method. The method can be used for the identification of the protein kinase phosphorylation specific sites, explore the protein phosphorylation rules under physiological and pathological states, further elaborate the nature of life and the disease pathogenesis and provide the important support for developing new drugs.

The invention discloses a train compartment air ventilation adjusting method and system based on microorganisms and a storage medium. A ventilation system is adjusted according to microorganism diffusion between measuring points, and microorganism pollution indexes of areas where passengers are located is reduced. The method has a guiding effect on air quality adjusting of railway trains. A mapping relationship between microbial pollution and atmospheric pollutant concentration is studied, a real-time problem of a microbial detection can be effectively solved, and real-time adjusting of the microbial pollution in train carriages is ensured.

The invention is a side leakage photon crystal optical fiber, and its structure has one or two light wave high-equivalent refractivity linear regions extending outward from core region to pass through coating region and being parallel with the fiber core and having cross section height between half wavelength and wavelength of transmitted light, and its preparing method comprises the three steps of material selecting and processing, precast bar making, charging and wire drawing on wire drawing tower; the transmission losses of two vertical polarizing state of its guide mode have obviously difference, and it can use as single polarizing optical fiber itself and because of having stronger rotation sensitivity, it is suitable to small-rotating angle optical fiber sensing, and can be used for making standardized and automated production of high quality optical fiber couplers.

Embodiments of the invention provide a pre-training model-based information extraction method, apparatus and device, and a storage medium. The method comprises the steps of obtaining a to-be-processedfinancial text; and extracting a target element from the financial text to be processed according to a preset information extraction model, wherein the information extraction model is obtained by training a context pre-training model. The embodiment of the specification can improve the prediction accuracy of the information extraction model.

The invention discloses a poultry food intake parallel anomaly detection method based on BIM and artificial intelligence. The poultry food intake parallel anomaly detection method comprises the stepsof: splicing and fusing poultry house images to obtain a poultry house panoramic image, and carrying out projection transformation on the poultry house panoramic image to a poultry house building information model; conducting poultry detection on the poultry house panoramic image; carrying out head key point detection on a region-of-interest image of each poultry in a period of time after the foodis fed; carrying out heat map superposition on the poultry head key point thermodynamic diagrams of which the poultry head key points fall into the a feeder region-of-interest within a period of timeafter the food is fed; inputting a superposition result into a poultry abnormal food intake analysis neural network to obtain a poultry abnormal food intake grade, and generating an abnormal food intake signal; and visualizing various kinds of information in the poultry house building information model. According to the poultry food intake parallel anomaly detection method, automatic poultry foodintake state monitoring is realized, the detection efficiency is high, and the detection accuracy is high.