31results about How to "Simplify the feature extraction process" patented technology

The invention discloses a traffic identification method based on a bag of word (BOW) model and statistic features. The method adopts the BOW model matched with a feature extraction method, trains collected network traffic features, and thus obtains a feature vector corresponding to each network category. For new network traffic, similarly, by extracting traffic features, utilizing the BOW module to obtain a corresponding feature vector, and then sequentially comparing with a feature vector of each network category which is previously established, the category corresponding to the feature vector with highest matching degree serves as a category tag of the new network traffic. The BOW method combines with an unsupervised k-means clustering method and a supervised k-nearest neighbor method, thereby being more suitable for multi-category classification. Due to the fact that the BOW model is not sensitive to space position, during feature extraction, arrangement according to time series of features is not required, and processing is convenient.

The invention discloses a face recognition method based on reference features. The method comprises the following steps that: scale invariant features and local binary pattern features of a face image to be recognized are extracted; a principal component analysis method is utilized for dimensionality reduction to obtain the image features of the face image to be recognized; the similarity of the image features to a cluster center is calculated by utilizing the obtained image features to obtain the reference features of the face image to be recognized; and the similarity of the reference features of the face image to be recognized and the reference features of training data concentration is calculated to obtain an analysis result. The reference features of the face image provided by the invention comprise texture information and structure information of the face image, so that the method provided by the invention can more comprehensively represent the face compared with the method in the prior art, which only represents the texture information or the structure information of the face. The process of feature extraction is simple and easy to realize; the recognition result is highly precise; high recognition rate of different facial gestures of the same person is realized.

The invention relates to a surface defect detection method, a terminal device and a storage medium, and the method comprises the steps: S1, collecting defect images of a detection object to form a training set, and marking the defect images in the training set; S2, constructing a defect detection model, and inputting the training set into the defect detection model for training to obtain a trained defect detection model; wherein the defect detection model is constructed based on a Faster R-CNN network, a feature extraction network of the defect detection model is a VGG-16 network, and the output of the third layer and the output of the fifth layer are superposed in the VGG-16 network; S3, inputting a defect image to be detected into the trained defect detection model to obtain a defect positioning frame and a defect type in the defect image; and S4, segmenting the defect according to the defect positioning frame in the defect image to be detected. Based on the Faster R-CNN network and the threshold segmentation method, defect types, positions and contours can be output only by inputting images in the detection process, and then end-to-end detection of defects is achieved.

The invention provides a face part identification method and device. The method comprises: obtaining a depth image; extracting image pixel characteristics in the depth image; inputting the image pixel characteristics into a face deep learning model to perform identification and classification; determining whether the classification of the image pixel characteristics matches existing face part labels in the face deep learning model; and if the classification of the image pixel characteristics matches existing face part labels in the face deep learning model, outputting the labels corresponding to the image pixel characteristics. According to the invention, a depth image characteristic extraction method is employed to guarantee extraction accuracy, and a deep learning model is employed to identify image pixel characteristics, and is capable of performing identification and classification on a plurality of face parts once.

The invention relates to an identity recognition method based on photoplethysmography. The identity recognition method comprises the following steps that S1, a data preprocessing program module preprocesses collected PPG signals; S2, a generalized S transformation program module performs generalized S transformation on the preprocessed PPG signal to obtain a PPG spectrum feature map; S3, a Getframe image processing module calls a Getframe function to snapshot the PPG spectrum feature map at each time point, and a continuous PPG spectrum trajectory feature map is obtained; and S4, the convolutional neural network performs feature extraction and feature classification on the PPG frequency spectrum track feature map to realize identity recognition. The method has the beneficial effects that the PPG signal is converted into the two-dimensional image from the original one-dimensional signal, so that identity recognition is carried out by adopting the convolutional neural network subsequently, the data features are easy to extract, the data feature extraction process is simple, and the reliability is high.

The invention provides a segmentation method for crop canopy images based on mean shift. The steps comprise firstly, re-sampling the crop canopy images in RGB color space, then, exchanging the crop canopy images in HSI space, next, indicating each pixel in the crop canopy images as a tetrad combined by four feature values, namely G-R, G-B, H, S, wherein R, G and B respectively indicate red weight, green weight and blue weight of pixel in the RGB space and H and S respectively indicate chroma and saturation of pixel in the HSI color space, then, employing mean shift algorithm to segment the crop canopy images into different types, finally, calculating the feature typical values of each type, if the first and the second components of the typical value are larger than zero, then, it is crops, if not, then, it is not crops. The invention has the advantages that parameters needed to be set is relatively less, the process of features extraction is simple, the algorithm is easy to be realized, and segmentation accuracy rate is high.

The invention belongs to the field of machine learning and pattern recognition, and particularly discloses a method for detecting man-machine mouse tracks on the basis of convolutional neural networks. The method includes preprocessing mouse track sampled data to obtain original characteristics such as coordinate and time characteristics, differential characteristics, speed characteristics, acceleration characteristics and direction characteristics with consistent lengths; carrying out standardized processing on the original characteristics, automatically extracting advanced semantic characteristics by the aid of the convolutional neural networks and carrying out training and prediction. The method has the advantages that characteristic extraction procedures can be simplified, the convolutional neural networks are excellent in generalization capacity, and machine attack means in man-machine verification products can be discriminated.

The invention provides a human body target identification method and apparatus. The method comprises the steps of obtaining a depth image; extracting image pixel characteristics in the depth image; inputting the image pixel characteristics into a human body deep learning model to perform identification and classification; judging whether categories of the image pixel characteristics are matched with existing human body part tags in the human body deep learning model or not; and if the categories of the image pixel characteristics are matched with the existing tags in the human body deep learning model, outputting the tags corresponding to the pixel characteristics. According to the method and the apparatus, the image pixel characteristics are identified by adopting the deep learning model, and human body target detection and identification are finished, so that the detection and identification process is simplified and the detection and identification efficiency is improved.

The invention provides a lithium battery life prediction method based on feature screening. The lithium battery life prediction method comprises the following steps: S1, dividing acquired lithium battery data into a training set and a test set, each of the training set and the test set comprising a plurality of samples; S2, performing pairwise non-repeated subtraction on the discharge SOCs in thefirst m charge-discharge cycles of the lithium battery samples in the training set and the test set; setting m to be larger than 20; S3, obtaining a variance of the SOC subtraction result of each lithium battery, and obtaining the features of each lithium battery; S4, inputting the features and the service life of the training samples in the training set into a Gaussian process regression model with ARD for model training; S5, visualizing the sparse feature weight of the trained model; and S6, inputting the features of the samples in the test set into the training model for life prediction. According to the invention, high correlation characteristics can be automatically extracted, so that the lithium battery life prediction is more accurate.

The invention pertains to the signal treatment and pattern recognition technique. Particularly speaking, the invention discloses an extraction method for wavelet features based on fuzzy wavelet-packet decomposition by taking stationary signals or non-stationary signals as signal samples, comprising the training process to signal samples, the category of which is marked, and the extraction process to new signal samples, the category of which are unknown; the training process is taken as main treatment process through the training process to find optimum wavelet decomposition Omega<*>, and based on the optimum wavelet decomposition Omega<*> to extract wavelet coefficients features with high identification performance; in the feature extraction process of the unknown category samples, the located wavelet coefficients are extracted as final features. By adopting the invention to treat the stationary signals and non-stationary signals (including violent changing signal) and extract wavelet coefficient features with strong identification performance, the distance of signals within the same category is made as small as possible, while, the distance of signals of different categories is made as large as possible; thereby, the classification of stationary signals and non-stationary signals are finally achieved.

The invention discloses a live pig walking stride frequency extraction method based on depth image skeleton end analysis. The extraction method includes the steps of removing backgrounds in depth camera images, extracting skeletons, pruning the skeletons and matching skeleton drawings, the ends of the pectoral appendage skeleton and the pelvic appendage skeleton of a live pig in all frames of images are extracted, the far side attribute and the near side attribute of the ends of the two skeletons are judged, a sine curve is fitted through change point sets of coordinate relations of the far sides and the near sides of the fore limbs and the posterior limbs of the live pig in the image frames of a live pig motion sequence, and then the live pig walking stride frequency characteristics are extracted. The extraction method has the advantages of long distance, non-invasion and automatic extraction based on computer vision; the feature extraction process is simple and easy to implement; the extraction method not only is suitable for collecting the live pig walking stride frequency features, but also can be used for detecting walking stride frequency of a horse, a bull, a sheep and other animals walking with the four limbs.

The invention discloses a time-frequency matrix dynamic selection based sound event classification method and system. The time-frequency matrix dynamic selection based sound event classification method comprises steps of collecting sound signal data in a set regional environment and preconditioning the sound signal data; generating a spectrogram for the preconditioned sound signals; zooming out the original spectrogram gradually to generate multiple time-frequency matrices different in size; obtaining similarity of each time-frequency matrix and the original spectrogram to find an optimal time-frequency matrix and converting the optimal time-frequency matrix to graph signals; extracting characteristic events from the graph signals; and sending the extracted characteristic events to a classifier to obtain a classification result of the sound events. The time-frequency matrix dynamic selection based sound event classification method has beneficial effects that the characteristic extraction process is simplified and a suitable dynamic threshold value is arranged to ensure completeness of the extracted characteristics. As a graph signal based method is adopted for similarity calculation of two images and a time-frequency matrix is selected dynamically for each sound signal, high-energy spectrum information of sound signals can be retained as far as possible while the calculation amount is reduced.

The invention provides a print-form Arabic alphabet identification method based on boundary characteristic, wherein the upside, downside, left and right boundaries is regarded as a wave, each boundary is regarded as an assembly of a series of wave elements, and the following boundary characteristics are extracted therefrom: the number of the component wave, the number of the zero line, the length of the first zero line in the right boundary, the length of the first zero line in the downside boundary, the length of the longest zero line in the upside boundary, the length of the longest zero line in the right boundary, the length of the longest zero line in the upside boundary, the length of the longest zero line in the downside boundary and the number of the positive line in the upside boundary, moreover, the height width ratio of the alphabets and the alphabet auxiliary part are involved as the identification characteristics, finally, each print-form Arabic alphabet is identified by using four decision trees according to the four kinds of formats of the alphabets-the independent, the beginning, the middle and the end.

The invention discloses a method and device for estimating the image blurring degree. The method includes the steps that an image is obtained, then the image is converted, and multi-scale representation of the image is obtained; error vectors between gradient normalization histograms and an original gradient normalization histogram of the image under all scales are calculated, and error vectors under all the scales are weighted and summed with reciprocals of quadratic sums of the error vectors; the estimated value of the blurring degree is obtained based on the error vectors weighted and summed under all the scales. According to the method and device for estimating the image blurring degree in the technical scheme in the embodiment, the good evaluating effect on blurring images with different kinds of content is achieved, and the evaluation result and the human-eye evaluation result have the high correlation.

The invention provides an on-line identification method of a compound pattern, which comprises the following steps of: firstly defining the characteristic of the compound pattern based on the formation of the element and the formation of the space relationship, wherein the characteristic is a set of combined characteristic vectors, the length of the characteristic vectors is the sum of the type of the component and the type of the space relationship, and each vector represents the number of the type of the element or the type of the space relationship in the formation of the compound pattern;designing a method for extracting the increment of the combined characteristic based on the definition of the combined characteristic vectors, passively updating the combined characteristic vectors along with the continuous pen input of the user, and performing the characteristic match only based on the updated vectors, so that the method simplifies the process for extracting the characteristic and guarantees the instantaneity of the characteristic match; and designing a space relationship erroneous judgment treating method in the process for the characteristic match, so that the method further guarantees the adaptability of the optional input of the user. The method has certain universality and expandability, is applied to the different professional fields, and realizes the intelligent human-machine interaction based on the hand painted draft on-line identification.

The invention provides a protein glycation locus identification method. The method comprises collecting protein glycation locus data, and extracting peptide chains which take lysine as the center fromthe protein glycation locus data to obtain a peptide chain sample set; encoding each amino acid of the peptide chain by adopting a single heat vector to obtain a peptide chain training set expressedby using the single heat vectors; training and producing an artificial peptide chain sample by using LSTM RNNs, and constructing an artificial peptide chain sample set; dividing each peptide chain inthe peptide chain sample set and the artificial peptide chain sample set into a series of biological words, and constructing features of peptide chains in the peptide chain sample set and the artificial peptide chain sample set through ProtVec based on the biological words; and obtaining a predictor based on CNN training, and identifying the protein glycation locus. According to the method of theinvention used to identify the protein glycation locus, feature extraction is less complex, and the accuracy of protein glycation locus identification is improved.

Disclosed is a registering method for optical and radar images: (1) a radar image is used as a reference image and an optical image is used as a to-be-registered image and downsampling is performed on the reference image and the to-be-registered image respectively so that images of no less than three layers and different resolutions are generated; (2) starting from the image of the first layer and low resolution, image transformation is performed on the image of each layer through use of mutual negative information values; (3) characteristic point sets of coordinates at which gradient magnitudes in the optical and the radar images exceed the gradient magnitude of a preset threshold value, are extracted respectively; (4) translational transformation parameters used in transformation of the image of the last layer in step (2) are used to transfer optical-image coordinate characteristic point sets extracted in the step (3) to radar-image coordinate characteristic point sets; (5) within the range of the transferred point sets, a target function is optimized and a translational transformation parameter corresponding to the maximum of the target function is selected as a fine registration parameter; (6) the fine registration parameter is used to perform transformation and resampling on the to-be-registered image so that a registered image is obtained.

The invention relates to a fingerprint feature extraction method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining a field fingerprint image; inputting the field fingerprint image into a fingerprint feature extraction model to obtain a feature extraction result output by the fingerprint feature extraction model, wherein the fingerprint feature extraction model is obtained by training based on a sample fingerprint image and detail point labels, wherein the fingerprint feature extraction model is used for extracting detail point information from the field fingerprint image and obtaining the feature extraction result based on the detail point information. According to the method and device, the electronic equipment and the storage medium provided by the embodiment of the invention, detail point information extraction is performed through the fingerprint feature extraction model, the feature extraction process can be simplified, and the model training speed and the feature extraction speed are increased while the accuracy is ensured.

The invention provides a local reflection symmetry axis extraction method in an image based on multi-instance subspace learning. The method comprises the following steps of 1, acquiring a local image multi-instance characteristic package of a natural image; 2, constructing a stepwise random projection tree; and 3, performing judgment analysis, namely analyzing the multi-instance characteristic package of the natural image of the to-be-extracted symmetry axis and a multi-instance characteristic package in a training data set by means of a subspace classifier, thereby obtaining an analysis result. The local reflection symmetry axis extraction method has advantages of simple and applicable characteristic extraction process, relatively high execution result recall rate and relatively high accuracy.

The invention discloses a face recognition method based on reference features. The method comprises the following steps that: scale invariant features and local binary pattern features of a face image to be recognized are extracted; a principal component analysis method is utilized for dimensionality reduction to obtain the image features of the face image to be recognized; the similarity of the image features to a cluster center is calculated by utilizing the obtained image features to obtain the reference features of the face image to be recognized; and the similarity of the reference features of the face image to be recognized and the reference features of training data concentration is calculated to obtain an analysis result. The reference features of the face image provided by the invention comprise texture information and structure information of the face image, so that the method provided by the invention can more comprehensively represent the face compared with the method in the prior art, which only represents the texture information or the structure information of the face. The process of feature extraction is simple and easy to realize; the recognition result is highly precise; high recognition rate of different facial gestures of the same person is realized.

The invention relates to a human body monitoring technology, in particular to a multi-lead physiological signal analysis method and device, which are used for reducing the analysis complexity of multi-lead physiological signals and reducing the hardware implementation cost. The method comprises the following steps of: converting the multi-lead physiological signal of the monitored object into two-dimensional time-frequency domain information from one-dimensional time domain information; since the two-dimensional time-frequency domain information can be regarded as a 'image-like', depth featuredata extraction and analysis processing can be carried out on the data frame by adopting an image algorithm in a subsequent process. Therefore, the current physiological state of the monitored objectis accurately determined, the analysis efficiency of physiological signals is improved, the early-stage working pressure of doctors can be reduced, the working efficiency of hospitals is improved, meanwhile, the feature extraction process can be simplified, the hardware implementation cost is reduced, and algorithm popularization is facilitated.

A method and apparatus for estimating the blurriness of an image are disclosed. In this method, an image is obtained, and then the image is transformed to obtain a multi-scale representation of the image. An error vector between the gradient normalized histogram at each scale and the original gradient normalized histogram of the image is calculated, and the error vector at each scale is weighted and summed by the inverse of their sum of squares. The ambiguity estimation value is obtained based on the weighted and summed error vectors at each scale. Using the solutions of the above embodiments, the evaluation effect on blurred images of various contents is good, and the evaluation results have a high correlation with the evaluation results of human eyes.

The invention discloses a traffic identification method based on a bag of word (BOW) model and statistic features. The method adopts the BOW model matched with a feature extraction method, trains collected network traffic features, and thus obtains a feature vector corresponding to each network category. For new network traffic, similarly, by extracting traffic features, utilizing the BOW module to obtain a corresponding feature vector, and then sequentially comparing with a feature vector of each network category which is previously established, the category corresponding to the feature vector with highest matching degree serves as a category tag of the new network traffic. The BOW method combines with an unsupervised k-means clustering method and a supervised k-nearest neighbor method, thereby being more suitable for multi-category classification. Due to the fact that the BOW model is not sensitive to space position, during feature extraction, arrangement according to time series of features is not required, and processing is convenient.

The present invention relates to a surface defect detection method, terminal equipment and storage medium. The method includes: S1: collecting defect images of detection objects to form a training set, and marking the defect images in the training set; S2: constructing defect images The detection model is trained by inputting the training set into the defect detection model to obtain the trained defect detection model; the defect detection model is constructed based on the Faster R-CNN network, and its feature extraction network is a VGG-16 network, and the VGG-16 network , superimpose the output of the third layer and the output of the fifth layer; S3: Input the defect image to be detected into the trained defect detection model, and obtain the defect location frame and defect type in the defect image; S4: According to the defect image to be detected The defect localization frame in the defect image is used to segment the defect. The invention is based on the Faster R-CNN network and the threshold segmentation method, so that the defect type, position and outline can be output only by inputting an image during the detection process, that is, the end-to-end detection of defects is realized.

The invention provides a segmentation method for crop canopy images based on mean shift. The steps comprise firstly, re-sampling the crop canopy images in RGB color space, then, exchanging the crop canopy images in HSI space, next, indicating each pixel in the crop canopy images as a tetrad combined by four feature values, namely G-R, G-B, H, S, wherein R, G and B respectively indicate red weight, green weight and blue weight of pixel in the RGB space and H and S respectively indicate chroma and saturation of pixel in the HSI color space, then, employing mean shift algorithm to segment the crop canopy images into different types, finally, calculating the feature typical values of each type, if the first and the second components of the typical value are larger than zero, then, it is crops, if not, then, it is not crops. The invention has the advantages that parameters needed to be set is relatively less, the process of features extraction is simple, the algorithm is easy to be realized, and segmentation accuracy rate is high.

The present invention provides a method for extracting local reflection symmetry axes in images based on multi-instance subspace learning. The method includes the following steps: step 1, obtaining local image multi-instance feature packages of natural images; step 2, constructing partial random projection tree; Step 3, decision analysis: analyze the multi-instance feature package of the natural image whose symmetry axis is to be extracted and the multi-instance feature package in the training data set with a subspace classifier to obtain the analysis result. The feature extraction process of the invention is simple and easy, and the recall rate and accuracy rate of the extraction result are high.

The invention discloses a live pig walking stride frequency extraction method based on depth image skeleton end analysis. The extraction method includes the steps of removing backgrounds in depth camera images, extracting skeletons, pruning the skeletons and matching skeleton drawings, the ends of the pectoral appendage skeleton and the pelvic appendage skeleton of a live pig in all frames of images are extracted, the far side attribute and the near side attribute of the ends of the two skeletons are judged, a sine curve is fitted through change point sets of coordinate relations of the far sides and the near sides of the fore limbs and the posterior limbs of the live pig in the image frames of a live pig motion sequence, and then the live pig walking stride frequency characteristics are extracted. The extraction method has the advantages of long distance, non-invasion and automatic extraction based on computer vision; the feature extraction process is simple and easy to implement; the extraction method not only is suitable for collecting the live pig walking stride frequency features, but also can be used for detecting walking stride frequency of a horse, a bull, a sheep and other animals walking with the four limbs.

The invention discloses a terminal model characteristic data cleaning system and cleaning method. The method comprises the following steps: S1, collecting message information generated when terminal equipment surfs the Internet and issuing the message information; S2, receiving the message information, obtaining features capable of representing the model of the terminal equipment from the messageinformation, and cleaning the features to form terminal model feature data; S3, discovering effective features in the terminal model feature data and determining the effectiveness of the effective features; S4, adjusting the terminal model characteristic data, and supplementing the terminal model characteristic data; S5, performing feature rule verification and feature validity verification on thefinally obtained terminal model feature data. According to the method, the feature extraction process is greatly simplified, excessive manual participation is avoided, manpower resources in enterprises are saved, and the feature data processing efficiency is remarkably improved.