41results about How to "Comprehensive feature extraction" patented technology

The invention discloses an SAR target identification method based on s CNN. The achieving steps are that 1. a target to be identified in each training image is subjected to multi-time random translation transformation, new samples are obtained, and the new samples are marked with labels of original images and are put into training samples in an expansion mode; 2. a convolutional neural network (CNN) structure is established in a caffe framework; 3. the training samples obtained after expansion are input into the CNN for training, and a trained network model is obtained; 4. a testing sample is subjected to multi-time translation expanding, and the testing sample obtained after expanding is obtained; and 5. the testing sample obtained after expanding is input into the trained CNN network model for testing, and the recognition rate is obtained. A target to be identified at any position of a sample image has the high recognition rate and stable performance, and the problem that according to an existing SAR target recognition method, influence from the position of the target to be recognized in the sample images is large is solved.

The invention provides an olfactory analog instrument and an on-site analysis method for the odor grade of a specific substance. The invention has the following characteristics: 1, a gas sensor array constant temperature operation room, a computer and an autoinjection system are integrated in a test box; 2, utilization of a sample bottle with a volume of 250 ml and 25 ml of a to-be-tested sample enables a gas-liquid ratio to be 9: 1; 3, parallel resistance wires respectively wind in a semi-circle manner to generate 45-W power, so the sample and headspace volatilized gas are heated to a temperature of 80 DEG C in only 8 min and maintained at the temperature for 30 min; 4, 25 ml of an aqueous ethanol solution with a concentration of 100 ppm is used to generate standard reference gas in the sample bottle with a volume of 250 ml, and correction on gas sensors are carried out based on the generated standard reference gas; 5, headspace sampling flow is 500 ml / min, sampling time is 30 s, and gas sensor response signals undergo lowpass filtering and dimensionality reduction pretreatment; and 6, a database is established, and the olfactory analog instrument carries out on-site detection and grade prediction and identification on the odor of specific substances consisting of an adhesive, petroleum wax, leather, glycerin and edible vegetable oil.

The invention discloses a cervical liquid-based cell slice quality detecting system and particularly relates to the field of artificial intelligence medical image processing. The system comprises a slice data processing module, a squamous cell quantity evaluation module, a fuzzy detection module, an abnormality detection module and an integrated module. An output end of the slice data processing module is connected to an input end of the squamous cell quantity evaluation module. An output end of the squamous cell quantity evaluation module is connected to an input end of the fuzzy detection module. An output end of the fuzzy detection module is connected to an input end of the abnormality detection module. An output end of the abnormality detection module is connected to an input end of the integrated module. The system is mainly for an x-slice image scanned by a scanner, and the detection content further includes the clarity and the number of squamous cells. Since the slice image includes the entire field of view of the slice, different situations may exist in different places in the field of view, and so the quality detecting system makes an objective judgment on the overall image with the combination of a slice thumbnail.

A quick ferrographic analysis method based on digital video includes the following steps: adopting a designed flow-controllable micro-channel device and a ferrographic microscope to be integrated into a whole; enabling a diluted oil sample to pass through a micro-channel directly; acquiring a clear abrasive particle image by regulating the focal length of the ferrographic microscope; sending the moving video of abrasive particles in the micro-channel to a computer through a digital camera; directly analyzing image frames through computer software to realize accurate positioning and characteristic identification of a plurality of characteristic abrasive particles to form a characteristic library. As abrasive particles tumble in the fluid, spatial characteristics of the abrasive particles from different perspectives can be acquired through characteristic identification of different frame images; through digital video image acquisition and fast computer automatic analysis, the advantages of an off-line analysis technology and an on-line analysis technology are combined, a novel method of acquiring abrasive particle characteristics on the basis of dynamic video is proposed, and real-time, continuous, high-accuracy and efficient on-line monitoring for the abrasion condition of the equipment is realized.

The invention discloses a database abnormality query detecting method based on target condition relevance rules. The method comprises a rule mining method and an abnormality detecting method, wherein the rule mining method comprises an initialization phase, a frequent-item generation phase and a target condition relevance rule generation phase; the abnormality detecting method adopts a rule bank obtained in the rule mining process to carry out abnormality detection; and just when the rule existing in the rule bank meets the rule support conditions, the operation is normal, otherwise, the operation is abnormal. For the invention, the rule mining efficiency, the detection efficiency and precision are high; the detection failure and the omission factor are low; and the universality and the expansibility are high.

The invention provides a cross-modal understanding and generating method and device based on a multi-modal pre-training model, and the method comprises the steps of determining to-be-processed multi-modal information which comprises an image, a text and an audio; and inputting the multi-modal information to the multi-modal pre-training model, learning the correlation of the multi-modal information to obtain a fusion representation of the multi-modal information, and inputting the fusion representation to the understanding and / or generating unit to execute a cross-modal understanding and generating task to obtain an understanding result and / or a generating result. According to the method and the device provided by the invention, understanding and generation are carried out by combining three modalities of an image, a text and an audio, so that full application of information is realized. Through combination of two tasks of cross-modal understanding and cross-modal generation, the multi-modal pre-training model can perform feature extraction and cross-modal association construction more comprehensively, so that the accuracy of cross-modal understanding and generation is further improved.

The invention relates to a method for detecting a computer generated image and a natural image based on wavelet transformation, which belongs to the technical field of image detection. The method comprises the following steps: firstly, performing color channel transformation to an image to be detected, transforming a RGB color image into an HSV color image; secondly, transforming all channels of the image further to obtain a multi-dimensional eigenvector including statistic matrix characteristics and high-frequency filter characteristics; thirdly, extracting characteristics by utilizing characteristics of the computer generated image and the natural images; fourthly, rapidly judging the truth of the image by a support vector machine (SVM), and detecting whether the image is the computer generated image or the natural image. The invention adopts a technology of combining the statistic matrix characteristics and the high-frequency filter characteristics, and has the characteristics of high precision, complete characteristics extraction, complete detection and the like, thereby greatly improving the detection precision.

The invention relates to the field of image detection technology, in particular to a head and shoulder region detection method and device. According to the method, a to-be-detected image is acquired;the to-be-detected image is detected based on a network model which completes training and contains a feature extraction network layer, a candidate box generation network layer and a target detectionnetwork layer, and a corresponding detection result is obtained, wherein the feature extraction network layer has the functions of extracting fusion features, reserving original feature information and adjusting model size; and when it is judged that a head and shoulder region exists in the to-be-detected image based on the detection result, the position of the head and shoulder region in the to-be-detected image is determined. By the adoption of the method, feature extraction on the to-be-detected original image is more comprehensive, therefore, the candidate box generation network layer cangenerate candidate boxes according to multi-scale features, and the precision of the detection result of the original image, which has poor shooting quality and is not clear, is guaranteed.

The invention discloses an expressway toll station traffic flow big data prediction method based on multi-target regression. The method comprises the following steps: 101, carrying out preprocessing operation on data; 102, marking the data; 103, performing feature engineering construction operation on the data; 104, constructing a multi-target regression model combining the specific characteristics of the target and the target correlation; and 105, predicting the traffic flow of the toll station every 20 minutes from 8 o'clock to 10 o'clock according to the historical traffic flow data of thetoll station, the weather data and other information through the established model. The method mainly preprocesses and analyzes historical traffic flow data, weather data and other information of a toll station to extract characteristics; a multi-target regression model combining target specific characteristics and target correlation is established, and the traffic flow from 8 o'clock to 10 o'clock per 20 minutes is predicted, so that a traffic management department can take measures in time by using big data to reduce the congestion of the toll station.

The invention discloses a method for constructing a similarity matrix in the ultrasound image Ncut segmentation process. The method includes the following steps of firstly, preprocessing an ultrasound image; secondly, conducting over-segmentation on the ultrasound image through a simple linear iterative clustering method to generate even-medium sub-areas with irregular borders; thirdly, enabling average gray level information of all the sub-areas to serve as one characteristic for constructing the Ncut similarity matrix, and calculating the texture characteristics of all the sub-areas through a gray level co-occurrence matrix at the same time, wherein the texture characteristics totally include 24 characteristics with six characteristics in each of the four directions, and the 24 data are combined to form a texture characteristic vector, namely, another characteristic for constructing the Ncut similarity matrix; fourthly, calculating the distance between the characteristics of every two adjacent sub-areas, and combining the sub-areas according to a certain proportion to form a new Ncut similarity matrix calculation formula. When the method is used for segmenting the ultrasound image based on Ncut clustering, a good segmentation result is obtained, the problems that the ultrasound image is high in noise and low in contrast ratio are solved, and the tumor areas and the background areas can be effectively separated.

The present invention belongs to the digital image processing technical field and discloses a selective extension-based fingerprint image matching method. According to the method, small fingerprint image matching is carried out by using small node and directional field information; fingerprint blocks with small nodes and fingerprint blocks without small nodes are considered differently through extending the functions of directional fields, so that feature extraction is realized; and matching scores are fused, so that the extracted features of small fingerprints can be fully applied. With the method adopted, more comprehensive feature extraction can be realized. Since the matching scores are fused, the extracted features of the small fingerprints can be fully applied, and accuracy can be improved. The method can accurately match the small nodes of fingerprints and can be used in an automatic fingerprint identification system.

The invention discloses a homology unit grouping method combining wavelet fuzzy entropy and GG fuzzy clustering, belonging to the field of power system unit homology identification. A phasor measurement device in a WAMS acquires post-fault system unit power angle information, the power angle swing curve is decomposed into global trend and detail information by using multi-scale wavelet decomposition, and the fuzzy entropy of wavelet coefficients of each scale is calculated. The fuzzy entropy of wavelet coefficients (including global feature and local feature) of each scale is used as the feature vector of a generator. The feature extraction is comprehensive and the calculation process is simple. Then the GG fuzzy clustering algorithm is used for homology clustering to improve the accuracyof grouping of homology units. The invention can obtain the grouping state of the system unit in real time, and provides necessary prerequisite for simplifying the electric network and determining thedisconnecting position.

The present invention discloses a method and a system for smog recognition based on a LBP (Local Binary Patterns) Gaussian pyramid. Suspected smog area images are subjected to graying and then twice Gaussian smoothing and sampling to obtain grey-scale maps with 1 / 4 and 1 / 16 sizes, and the grey-scale maps with 1 / 4 and 1 / 16 sizes are combined with an original image grey-scale maps to form three layers of pyramid images; LBP operators of P being equal to 8 and R being equal to 1 are employed for the three layers of Gaussian pyramid grey-scale maps to calculate and obtain binary system LBP codes of the three layers of Gaussian pyramid grey-scale maps, a rotation invariant mode and an equivalent mode are employed to perform dimension reduction of each layer of LBP codes, nine types of LBP codemodes are obtained, and statistics of the number of each type of LBP codes are employed to take the number of each type of LBP codes as one feature value; and AdaBoost input vectors are formed by employing 27 feature values of the three layers of LBP Gaussian pyramids for discrimination of smog and false smog interference. The method provided by the invention has good robustness and high recognition rate.

The invention discloses a laser point cloud three-dimensional target detection model and method for a complex traffic scene, a three-dimensional encoder in the model is beneficial to the detection of long-distance targets and small targets, and sparse convolution and sub-manifold convolution can greatly improve the coding efficiency of voxel features. The residual structure of the two-dimensional encoder can keep more complete information, detection of a long-distance target and a small target is facilitated, meanwhile, the network is easier to optimize, feature extraction and receptive field expansion can be carried out on an original scale and a self-calibration scale through self-calibration convolution, more complete and rich features can be extracted. Useful feature expression is enhanced in the space direction and the channel direction through space attention and channel attention, and useless information is inhibited. The final detection precision of vehicles is 81.88%, the final detection precision of pedestrians is 47.82%, the final detection precision of riders is 69.81%, the average precision is 66.25%, the average precision is 9.9% higher than that of an existing VOXEL RCNN algorithm, 13.8 FPS is achieved on RTX 2080Ti display, and the detection precision and speed can meet the sensing requirements of intelligent vehicles in complex traffic environments.

The invention relates to the field of semantic analysis, and discloses a text classification method, which comprises the following steps of: performing category label marking on each text in a text set to obtain a target label set of the text set; performing text splicing processing on the text set and the target label set to obtain a sample sequence set; performing iterative training based on neural feature fusion extraction on a pre-constructed text classification model by using the sample sequence set until the text classification model converges, and obtaining a trained text classification model; and when a to-be-classified text is received, carrying out word segmentation and label splicing on the to-be-classified text to obtain a to-be-classified text sequence, and classifying the to-be-classified text sequence by utilizing the trained text classification model to obtain a classification result. The invention also relates to blockchain technology, and the text set can be stored in a blockchain node. The invention further provides a text classification device, electronic equipment and a storage medium. According to the method, the accuracy of text classification can be improved.

The invention discloses an off-duty detection method and device and a storage medium. The off-duty detection method comprises the steps of obtaining a picture of a human body image containing a to-be-detected object; utilizing a feature extraction model based on machine learning training to determine position information of the human body image in the picture; and judging whether the to-be-detected object leaves the post or not according to the position information. The off-duty detection method provided by the invention is relatively comprehensive in identification, high in accuracy and highin identification speed, and has relatively low time delay, so that real-time monitoring is ensured. In addition, a feature extraction model based on machine learning training is adopted in the scheme; therefore, compared with a traditional detection method, whether a plurality of employees are on duty or not can be detected only by using a single camera, if a plurality of employees are in one camera, the employees in the range only need to mark out the position range of the employees when the position range is divided, the cost is low, the use is convenient, and no complex wearable equipmentis used.

The invention provides a pig body size and weight estimation method based on deep learning, and relates to the technical field of deep learning. According to the method, the convolutional neural network is used for predicting the weight of the pig, related features are obtained through learning of the convolutional neural network, feature engineering extraction does not need to be constructed, so that the extracted features are more comprehensive, and the convolutional neural network is superior to a linear model in noise data processing and data nonlinearity problems; and a pig picture is shot by a universal 2d color camera, the equipment price is low, and the cost is low when the technical scheme is implemented.

The invention discloses a medicine quality control analysis method and device based on machine learning, equipment and a medium, and relates to the technical field of information. According to the method, the feature space of a machine learning model can be expanded, and disease attribute features and medicine attribute features are introduced, so that the quality control analysis precision of themodel is improved. The method comprises the following steps: determining an identification feature vector corresponding to disease identification information and medicine identification information;determining an attribute feature vector commonly corresponding to the disease attribute information and the medicine attribute information; and determining a quality control analysis result of the medicine in the patient medical record according to the identification feature vector and the attribute feature vector. The invention relates to a machine learning technology in artificial intelligence,is suitable for quality control analysis of medicines and is also suitable for the field of smart medical treatment, so that the construction of smart cities can be further promoted. In addition, theinvention also relates to a blockchain technology.

The invention relates to an image recognition method based on VGG16 on insect taxonomy. The method comprises the following steps: S1, establishing an image data set; S2, processing insect images of the image data set to obtain a training data set; S3, training the training data set by using a VGG16 model; S4, extracting a part of the image from the image data set as a reference image and a to-be-identified image, and performing corner detection to correct the reference image; S5, processing the image to be identified and the corrected reference image, inputting the processed image to be identified and the corrected reference image into the trained VGG16 model, and extracting image features; S6, visualizing the extracted image features to obtain a feature map; and S7, calculating the feature map image similarity SSIM of the to-be-identified image and all reference images under each type of insect mesh-level order, solving a mean value, and classifying the to-be-identified image to the type with the maximum mean value to serve as the mesh-level order to which the to-be-identified image belongs. According to the invention, the insect classification accuracy and efficiency are improved.

The invention discloses a data preservation message request processing method. The data preservation message request processing method comprises receiving an identity authentication request sent by aclient SDK, verifying user information carried by the identity authentication request, and generating and returning an authorization token and a private key; receiving a data preservation message sentby the client, filtering the message, and releasing the message passing the filtering process; carrying out message header compliance verification on the released message, carrying out tamper-proofing verification on the client SDK, and carrying out legality verification on the token; analyzing the message header and the service content of the message, and obtaininng the verification value carried by the message header and service content structural data; calculating the verification value of the obtained service content structural data, comparing the verification value with the obtained verification value carried by the message header, and carrying out message tamper-proofing verification; and decrypting the service content structural data according to the private key, and obtaining decrypted preservation data. According to the invention, the problem that in the prior art, the message processing efficiency is low, and the message filtering accuracy is low is solved.

The invention discloses a CNN-based SAR target recognition method, the implementation steps of which are as follows: 1. Perform multiple random translation transformations on the target to be recognized in each training image to obtain new samples, and mark these new samples with original The label of the image is expanded into the training sample; 2. Construct the CNN structure of the convolutional neural network in the caffe architecture; 3. Input the expanded training sample into the CNN for training to obtain the trained network model; 4. Test the sample Multiple times of translation and expansion to obtain the expanded test samples; 5. Input the expanded test samples into the trained CNN network model for testing to obtain its recognition rate. The invention has high recognition rate and stable performance for the target to be recognized at any position of the sample image, and solves the problem that the existing SAR target recognition method is greatly affected by the position of the target to be recognized in the sample image.

The invention discloses a database abnormality query detecting method based on target condition relevance rules. The method comprises a rule mining method and an abnormality detecting method, wherein the rule mining method comprises an initialization phase, a frequent-item generation phase and a target condition relevance rule generation phase; the abnormality detecting method adopts a rule bank obtained in the rule mining process to carry out abnormality detection; and just when the rule existing in the rule bank meets the rule support conditions, the operation is normal, otherwise, the operation is abnormal. For the invention, the rule mining efficiency, the detection efficiency and precision are high; the detection failure and the omission factor are low; and the universality and the expansibility are high.

The present invention - an olfactory simulation instrument and a method for on-site analysis of the gas (smell) taste level of specific substances, one of the characteristics, the test box integrates a gas sensor array constant temperature studio, a computer and an automatic sampling system; the second characteristic, 250 ml The sample bottle and 25ml of the sample to be tested make the gas-liquid ratio reach 9:1; the third feature is that the parallel resistance wire is wound in half a circle to generate 45 watts of power, and it only takes 8 minutes to heat the sample and headspace volatile gas to 80°C, and the constant temperature is 30°C. Minutes; feature four, use 25 ml of 100ppm ethanol aqueous solution to generate a standard reference gas in a 250 ml sample bottle, based on which the gas sensor is calibrated in time; feature five, the headspace sampling flow rate is 500 ml / min, and the sampling time is 30 seconds. The response signal of the gas sensor is preprocessed by low-pass filtering and dimension reduction; the sixth feature is to establish a database, and the olfactory simulation instrument can conduct on-site detection and grade prediction and identification.

The invention provides a fault diagnosis method based on Transform, and belongs to the technical field of fault diagnosis. According to the method, multiple layers of Transform Encoder are adopted as a feature extraction module, Dense connection is added between Encoder layers to enhance the model feature multiplexing capability, a Dropout layer is added in front of the feature extraction module to improve the generalization capability of the model, a multi-channel convolutional neural network layer is added to generate a sample matrix, and a full connection layer is adopted to perform fault classification. According to the method, a Transform Encoder structure is adopted and improved, the Transform Encoder structure is applied to fault diagnosis of mechanical equipment, time sequence features and global features between vibration signals in a long time can be well extracted, and a more accurate fault relation is obtained.

The invention discloses a data preservation message request processing method. The data preservation message request processing method comprises receiving an identity authentication request sent by aclient SDK, verifying user information carried by the identity authentication request, and generating and returning an authorization token and a private key; receiving a data preservation message sentby the client, filtering the message, and releasing the message passing the filtering process; carrying out message header compliance verification on the released message, carrying out tamper-proofing verification on the client SDK, and carrying out legality verification on the token; analyzing the message header and the service content of the message, and obtaininng the verification value carried by the message header and service content structural data; calculating the verification value of the obtained service content structural data, comparing the verification value with the obtained verification value carried by the message header, and carrying out message tamper-proofing verification; and decrypting the service content structural data according to the private key, and obtaining decrypted preservation data. According to the invention, the problem that in the prior art, the message processing efficiency is low, and the message filtering accuracy is low is solved.

The invention discloses a fault diagnosis method of abnormal vibration, of a rotor, caused by a floating oil retainer. The fault diagnosis method of abnormal vibration, of a rotor, caused by a floating oil retainer includes the steps: importing shafting TSI data into a fault diagnosis system; setting a threshold, utilizing a spectrum analysis method, determining whether the property of shafting abnormity vibration belongs to common forced vibration, extracting typical characteristics of a fault, analyzing operating parameters of a set, and eliminating the influence of load, thermal deformation and field current; and with combination of shafting structure analysis structure, determining whether abnormal vibration of the rotor is caused by a floating oil retainer, and then obtaining a diagnosis result. The fault diagnosis method of abnormal vibration, of a rotor, caused by a floating oil retainer has the advantages of being comprehensive in the considered extracted characteristics, being suitable for field fault diagnosis, being quick and accurate in result diagnosis, and having practicality and operability.

The invention discloses a rotating machinery online intelligent fault diagnosis method and system for lifelong learning. The method comprises the following steps: building an initial intelligent diagnosis model through collected historical fault data, and carrying out the online updating of the intelligent diagnosis model through a sample incremental learning mode when new fault sample data is collected step by step, and when new fault mode data is collected step by step, updating the intelligent diagnosis model on line in a mode of mode incremental learning, collecting a data sample of the rotating machine to be diagnosed, inputting the data sample into the intelligent diagnosis model, and diagnosing the fault of the rotating machine. According to the method, lifetime learning of the intelligent diagnosis model is realized through initial model training, sample incremental learning and mode incremental learning, data stream information can be effectively processed, storage consumption of data is greatly reduced, and practical engineering application is facilitated.

The invention belongs to the technical field of image classification, and particularly relates to an EGC image classification method based on CNN + SVM, and the method comprises the following steps: converting all data into a grey-scale map, omitting the color information of a sample, increasing the contrast of an image, and carrying out the normalization processing of the data converted into thegrey-scale map, thereby obtaining an original data set; dividing the data set to obtain a training set and a test set; extracting data features in two modes of automatic extraction and manual extraction; and performing weighted summation on the classification results of the CNN and SVM algorithm models to obtain a final classification result. According to the method, a traditional classification method SVM and a deep learning method CNN are combined, feature extraction is carried out on the EGC data in a mode of combining manual features and automatic features, the interpretability of the model is improved through the use of the manual features, the potential safety hazard of the model is reduced, and the model is helped to have higher accuracy compared with a traditional method through the addition of the automatic features. The method is used for classifying the EGC images.

The invention discloses a method for constructing a similarity matrix in the ultrasound image Ncut segmentation process. The method includes the following steps of firstly, preprocessing an ultrasound image; secondly, conducting over-segmentation on the ultrasound image through a simple linear iterative clustering method to generate even-medium sub-areas with irregular borders; thirdly, enabling average gray level information of all the sub-areas to serve as one characteristic for constructing the Ncut similarity matrix, and calculating the texture characteristics of all the sub-areas through a gray level co-occurrence matrix at the same time, wherein the texture characteristics totally include 24 characteristics with six characteristics in each of the four directions, and the 24 data are combined to form a texture characteristic vector, namely, another characteristic for constructing the Ncut similarity matrix; fourthly, calculating the distance between the characteristics of every two adjacent sub-areas, and combining the sub-areas according to a certain proportion to form a new Ncut similarity matrix calculation formula. When the method is used for segmenting the ultrasound image based on Ncut clustering, a good segmentation result is obtained, the problems that the ultrasound image is high in noise and low in contrast ratio are solved, and the tumor areas and the background areas can be effectively separated.