78results about How to "Accurate feature extraction" patented technology

The invention discloses a monocular vision AGV (Automated Guided Vehicle) accurate positioning method and system based on multi-window real-time range finding. The method comprises according to a camera slanting installation mode, calibrating and measuring camera parameters and establishing a visual system real-time measuring model; setting a circular color lump on the ground to be used as the reference substance for parking positioning, identifying the circular color lump through the efficient algorithm of the visual system, and accurately extracting central position information; and in a view filed, setting a plurality of windows to process images a far-end window is used for an AGV to predetermine ground information so as to gradually decelerate; an intermediate window is used as a coarse positioning window, and used for adjusting poses; and a near-end window is used for accurate range finding and parking. The method allows an AGV to sense depth information, and has the advantages of high characteristic recognition rate, excellent arithmetic instantaneity, low cost and great extendibility. The horizontal distance deviation of AGV parking is stabilized at +-1 mm, the angle deviation is stabilized at +-1 DEG, and the parking error is stabilized at +- 2mm.

The invention provides a vehicle license character feature extracting and classifying method using projection symmetry as the precondition for judgment. The method comprises the following steps: firstly extracting character feature to 26 letters and 10 numbers which may appear in vehicle license characters by using projection symmetry as precondition to divide into four classes, namely vertical projection class, horizontal projection class, central point projection class and dissymmetric property class and realize the coarse classifications of vehicle license characters; and fining, performing normalization transformation, then extracting features of points and rings and completing the fine classification of vehicle license characters. The invention combines the projection symmetry with the feature extraction method of points and rings to set a vehicle license character classifier, thus laying the foundation of realizing the identification of vehicle license characters finally. The method has better identification effect on confusable characters, such as '0' and 'D', '8' and 'B', '7' and 'T' and the like, thus increasing the identification speed and accuracy of vehicle license characters.

The invention provides a method for identifying a hyperspectral material. First a standard light source spectrum database is established, a spectral material standard database is established, and standard light source spectra are obtained through a difference method and a control variable method; spectral data information is shot for a material needing to be identified; through preprocessing and intrinsic material information extraction, spectral characteristics of the material are obtained, and then the spectral characteristics are matched with data in the spectral material standard database, so as to perform high-precision identification of the material. According to the method, the interference to spectral information of the material caused by complex factors such as different illumination conditions and material shape changes and the like can be resisted, the problems of metamerism and homogeneity and heterochrome which cannot be solved through conventional red green blue (RGB) and RGBD cameras are solved, and the identification precision rate is high.

The invention discloses an image multi-subtitle automatic generation method based on a multiscale hierarchical residual network, and adopts an improved funnel network to capture multiscale target information. Firstly, when a funnel framework network is constructed, a densely connected polymerization residual block is put forward, and residual LSTM (Long Short Term Memory) is further put forward inorder to solve the problems of gradient vanishing and gradient explosion. By use of the method, high experiment performance is obtained, and the method has an obvious advantage on multi-subtitle taskacquisition.

The invention discloses an analysis method for recognizing attack behaviors of group-housed pigs through employing the machine vision technology. The method comprises the steps: extracting an attack key frame sequence from a downward view group-housed pig video, and locating an attacking pig; taking the attacking pig as the whole body for the extraction of the acceleration characteristics; carrying out the training of the acceleration data, obtaining an acceleration threshold value, and dividing the key frame into a high-level frame, a middle-level frame and a non-attack frame according to thethreshold value; finally setting a minimum unit of the attack recognition, and classifying the group-housed pigs into high-level, middle-level and non-attack group-housed pigs according to the proportion of the attack frame in the unit. The method is used for the recognition of attack behaviors of the group-housed pigs through the machine vision technology, does not causes any impact on the group-housed pigs, provides a theoretical basis for the exploration of the attack rule, the evaluation of the damage level and the determination of manual intervention, and also provides reference for thedetection of abnormal behaviors of other livestock based on the accelerated movement.

The invention belongs to the field of image classification, and relates to a super-large image classification method based on a graph neural network. Different sub-image screening methods can be adopted for different super-large images, and the feature extraction network is further adjusted, so that feature extraction of the sub-images is more accurate; a super-large image is constructed into image data; a differentiable pooling operation is introduced into a traditional graph convolutional neural network; and global information of the super-large image can be mined, feature information of a hidden layer can be mined in the training process through micro-pooling operation, the relevance of all the sub-images in the feature space is fully analyzed, and the super-large image can be classified more accurately.

The invention relates to an image recognition method and device, computer equipment and a storage medium. The method comprises: acquiring a to-be-processed image through the computer equipment; carrying out feature extraction on the to-be-processed image by adopting a preset identification model to obtain an identification vector, wherein the recognition model is a model obtained by training by adopting an attention mechanism and adopting a dense loss function, and the recognition vector is used for representing a plurality of local features of the to-be-processed image; and performing image recognition on the recognition vector to obtain a recognition result. With the adoption of the method, the accuracy of image recognition under the conditions of shielding or large-angle shooting and the like is greatly improved.

The invention discloses a feature extraction method for motor imagery electroencephalography signals. The method includes the steps that the optimal time period for feature extraction of motor imageryelectroencephalography signals is determined according to the average power spectrum, then, four-layer double-tree complex wavelet decomposition is performed on the motor imagery electroencephalography signals within the time period, signal reconstruction is performed with the complex wavelet coefficient of each sub-band, and the average energy feature of data in the optimal time period of the reconstructed signals is calculated to serve as the time-frequency feature of the motor imagery electroencephalography signals; an IL-MVU algorithm is proposed to perform dimensionality reduction on thedata in the optimal time period of the reconstructed electroencephalography signals, low-dimensional vectors obtained after dimensionality reduction are taken as nonlinear features of the motor imagery electroencephalography signals, and finally standardization and feature fusion are performed on the time-frequency feature and nonlinear features of the motor imagery electroencephalography signalsin the optimal time period to obtain feature vectors of the motor imagery electroencephalography signals. The method greatly reduces the time consumption of the algorithm and improves the classification accuracy of the MI-EEG signals.

The invention provides a vibration type flow meter characteristic signal extraction method based on improved assemble average empirical mode decomposition. The method includes the steps of conducting end continuation processing on a collected vibration type flow meter flow vibration signal through a waveform-matched self-adaption end continuation method, conducting envelope line fitting on the collected vibration signal through a cubic B-spline method, conducting MEEMD decomposition to obtain a plurality of IMF components, conducting relevance analysis on the IMF components and the original signal, selecting the useful IMF components, conducting HHT conversion on the IMF components, and obtaining the Hilbert time-frequency spectrum and the marginal spectrum of the flow signal, wherein the Hilbert time-frequency spectrum and the marginal spectrum are the signal characteristics of the vibration type flow meter flow vibration signal. The method is suitable for accurately and rapidly metering the pipe network fluid flow in the industrial field.

The invention provides a face image processing method, which comprises the following steps that: a first face image and a second face image of the face of the same person are received, wherein the first face image comprises a first face part and a first background part, and the second face image comprises a second face part and a second background part; the first face image and the second face image are stored in an image database; the first face part is extracted from the first face image, and the second face part is extracted from the second face image; a first face feature vector of the first face part and a second face feature vector of the second face part are extracted; and a face owner of the face contained in the first face image and the second face image is judged according to the first face feature vector and the second face feature vector. The pertinence and the accuracy of the face recognition and detection are effectively enhanced.

The invention discloses a three-dimensional point cloud head posture estimation system and method based on ordered regression and soft labels. The system comprises a feature learning network module which is used for the layered feature extraction of point cloud data; the prediction network module is used for mapping the features obtained by the feature learning network module to a head attitude angle to obtain an angle prediction value, and substituting the angle prediction value and the head attitude angle serving as a label into a first loss function; the sorting network module is used for carrying out dimension division on the head attitude angle to form a plurality of subtasks, generating a soft label according to the relationship between the head attitude angle serving as a label andthe subtasks, carrying out value prediction on the features obtained by the feature learning network module, and substituting the value prediction of the point cloud data and the soft label into a second loss function; network updating module. The loss of the sorting network module and the loss of the prediction network module are combined to introduce the sorting network so as to guide the learning of the prediction network, so that the feature extraction is more accurate, and the precision of the prediction network is improved.

The invention provides an Open Pose-based monocular camera gesture language recognition method comprising the steps of collecting video data of gesture language of a presenter with a video camera; inputting the collected video data to an OpenPose system and primarily extracting three-dimensional characteristic data containing x-axis coordinates, y-axis coordinates and the confidence coefficient; selecting primarily extracted characteristic points, rebuilding a coordinate system with the neck as the origin, performing normalization in the x-axis direction and the y-axis direction to obtain final characteristic data; scanning the characteristic data by using three different granularities to obtain expansion characteristic data; inputting the expansion characteristic data to a deep forest model for multi-layer meaning recognition and obtaining recognition results of final meaning through an extreme classifier from the output of the last layer. The method has the capability of monocular vision recognition of gesture language, does not need big sample data and has the advantages of accurate characteristic extraction, simple process and high meaning recognition accuracy.

The invention discloses a model training method and a device, a point cloud missing completion method and a device, electronic equipment and a computer readable storage medium. The model training method comprises the steps of obtaining training missing point cloud data; inputting the training missing point cloud data into an initial model to obtain training repair point cloud data, and adjusting parameters of the initial model based on the training repair point cloud data and original point cloud data corresponding to the training missing point cloud data; if it is detected that the training completion condition is met, determining that the initial model is a point cloud completion model; wherein the initial model comprises a target reconstruction network and an initial generation network, the target reconstruction network comprises a target coding network, the target coding network uses training missing point cloud data to carry out comparative learning, the training missing point cloud data is input into the target coding network to obtain input features, the input features are input into the initial generation network to obtain missing point cloud data, the missing point cloud data is used for generating training repair point cloud data; and the accuracy of the processed point cloud data after completion processing is improved.

The invention belongs to the technical field of digital image processing, and discloses a dual-Gabor palm print ROI matching method based on specific expanded eight neighborhoods. The dual-Gabor palm print ROI matching method comprises inputting an ROI image of a palm print to be matched and an ROI image of a template palm print, and carting out dual-Gabor filtering; carrying out orientation coding processing and obtaining coded images; carrying out normalized coding on the variation degree of the coded images, and screening out a point the variation degree of which is highest; carrying out image pyramid calculation on the two coded images to obtain 1, 1/2 scale transformation images; obtaining the offset of a sampling point of the image in each scale, and finding out the position of the sampling point after offset; obtaining a preliminary matching score by calculating relative positions between the sampling points and the relative positions between the sampling points after offset; calculating the overlap ratio of the two palm prints which are subjected to variation degree normalized coding, and obtaining a final matching score by combining with the preliminary matching score; and judging whether the matching is true by means of a set fixed threshold. The dual-Gabor palm print ROI matching method based on specific expanded eight neighborhoods can carry put palm print image ROI matching accurately.

To reduce the complexity of a face recognition algorithm, the invention provides a quick face recognition system based on personalized features especially lip images. The system reads face image dataof a person. Lip images at the same angle are collected at first, and then identification is conducted on the basis of a lip in the lip images and image data of other parts.

The invention discloses a rock image retrieval method which comprises: acquiring image data in real time, and inputting the image data into a trained deep learning network model to obtain a retrievalresult graph. The training process of the deep learning network model comprises: constructing a rock image data set by utilizing collected image data; inputting the data set into a network, and enabling the network to actively convert feature mapping in space after the data set is processed by a space transfer module; and inputting the processed data into a multi-granularity network, calculating atotal loss function and an mAP value of the model, and after multiple calculations, when the loss function tends to be stable and the mAP value reaches a peak value, completing training of the deep learning network model. The situation that only representation is used for classifying the rock images is avoided, meanwhile, the fine-grained characteristics of the rock images are extracted more accurately, and the retrieval accuracy of the rock images can be improved under the conditions that sundries are shielded, the number of samples is small, the quality is low, and information is lost.

The invention relates to a micro-expression recognition financial risk control method. The method comprises: obtaining and preprocessing image frames of a video; detecting the preprocessed image frameto obtain a micro-expression frame sequence in the image frame; extracting the detected micro-expression frame sequence to obtain micro-expression features; inputting the obtained micro-expression features into a trained convolutional neural network for classification; and performing psychological analysis on a classification result of the convolutional neural network, and estimating a fraud riskvalue. The invention further relates to a micro-expression recognition financial risk control system. According to the method, the potential fraud risk of the lending user is judged in a mode of combining micro-expression recognition and psychology, a more accurate financial fraud recognition method can be provided for a cash lending company, and the economic loss of the cash lending company is reduced.

The embodiment of the invention provides a banknote authenticity identification method and device. The method comprises the steps that the characteristic area including the denomination number is cutfrom the banknote grayscale image obtained through infrared scanning; the boundary information of the included denomination number in the characteristic area is determined; the projection informationof the characteristic area is obtained according to the boundary information; and the banknote authenticity is determined according to the projection information. With application of the technical scheme, characteristic extraction of the characteristic area can be accurately performed through the determined boundary information and then the banknote authenticity can be determined according to theprojection information of the boundary information so that the accuracy of banknote authenticity identification can be greatly enhanced.