32results about How to "Feature extraction is fast" patented technology

The invention discloses a low-resolution multi-spectral palm print and palm vein real-time identity recognition method and system. Palm images are collected by the system under the condition of five spectrums, and complementarity of multi-spectrum image information is fully utilized to improve the system recognition rate; meanwhile, palm vein information is collected under the condition of near infrared spectrums so that the system can have the living body detection ability and the counterfeit attack preventing ability of the system can be improved; characteristic extraction speed and other postprocessing speed are improved through the down sampling technology based on bicubic interpolation, and storage space of a characteristic template is saved; characteristic extraction is carried out through a multi-scale multi-directional filter, the influence of lighting changes on characteristic extraction is reduced, and the robustness of the system is improved; a characteristic matrix is coded through a hash table, and system matching speed is further improved; the recognition rate of the system is further improved through the unique fraction-level multi-spectral characteristic fusion method. The system has the advantages of being high in resolution ratio, high in recognition speed, good in stability and expansibility, resistant to counterfeit attack and the like.

The invention brings forward an aerial photography image and geographical data splicing method based on ORB feature matching. The method comprises the following steps: step 1, reading images to be spliced and geographical information; step 2, distributing tasks to processors; step 3, extracting ORB features of the images to be spliced; step 4, performing initialization estimation on parameters of a camera shooting the images to be spliced and solving a rotation matrix; step 5, improving estimation precision by use of bundle adjustment; step 6, performing initial splicing on the images to be spliced; step 7, performing brightness increment compensation on the images after the splicing, and based on multiband fusion of an image pyramid, performing conditional interpolation fusion on the geographical information after the splicing to obtain splicing result images; and step 8, and performing primary integrated splicing fusion on the splicing result images processed through all threads of each processor, and performing merging to obtain a final result image.

The invention provides a remote sensing image change detection method and system based on a U-net improved algorithm, and the method comprises: carrying out the preprocessing of a remote sensing image, and then integrating the preprocessed remote sensing images of different time phases to obtain an input image; performing down-sampling encoding and up-sampling decoding on the input image based onan improved U-net network; and finally, outputting a binary image whether the image is changed or not according to an analysis result. According to the invention, a network structure in a semantic segmentation direction is applied to the field of change detection, and a residual learning mechanism is introduced, so that the encoder can quickly converge and deepen the number of network layers; andmeanwhile, ASPP is used for enhancing the perception capability of the network to image features, so that the algorithm can be kept at a relatively high level in precision and efficiency, and has relatively high robustness. The method and the system are suitable for the field of change detection of remote sensing images, can be popularized to other fields, and have important significance.

The invention discloses an unmanned autonomous working platform based on an all-weather unknown environment, and belongs to the field of artificial intelligence and visual navigation. The platform comprises five modules: a stereoscopic vision positioning module, an infrared visible light fusion module, an image recognition module, a map construction module and a loop-back and return detection module. The visual positioning module and the image recognition module share a graph convolutional neural network framework, the visual positioning module selects key frames to perform feature matching and visual positioning, the image recognition module performs semantic classification on a point cloud local map, and the map construction module performs point cloud splicing to form a global depth dense semantic map. The deep neural network is introduced to improve the feature extraction effect and save the extraction time. Monocular vision distance measurement is adopted, so that the multi-parallax registration time is saved. Multi-spectral fusion of key frame images is carried out, all-weather efficient work is achieved, and the detection rate of shielded targets is increased.

The invention provides a method which can intelligently recognize Hong Kong dollar and quickly recognize Hong Kong dollar versions and relates to a Hong Kong dollar texture image version classifying method based on a gray-level co-occurrence matrix. A sample base is formed by extracting all versions of sample Hong Kong dollars, gray-level co-occurrence matrix characteristics of sample Hong Kong dollar images are extracted, sample is trained by a BP neural network, a Hong Kong dollar version classifying model is obtained, parameters are transplanted into a currency counting and detecting machine, a bill sorting machine and a money depositing and withdrawing machine, texture characteristics parameters, obtained through the gray-level co-occurrence matrix, of paper money to be recognized are inputted into a multispectral currency counting machine, the bill sorting machine and the money depositing and withdrawing machine, proof coin types can be calculated out according to a Hong Kong dollar version classifying model, and namely Hong Kong dollar versions can be judged.

The invention provides a human face recognition method and device based on a residual-quantized convolutional neural network, in order to provide a human face recognition method and device capable ofperforming large-scale human face recognition and reducing the amount of calculation, thereby reducing hardware requirements and reducing training time. The human face recognition method comprises thefollowing steps: step S1, constructing a convolutional neural network model and performing training; step S2, preprocessing a target image and performing preprocessing on a to-be-determined image; step S3, sequentially inputting the preprocessed to-be-determined image and the preprocessed target image to a feature extraction model to obtain a to-be-determined feature vector and a target feature vector; step S4, determining the consistent human face image according to the target feature vector and the to-be-determined vector, wherein the step S1 comprises a step of setting a predetermined layer to be a quantized layer, performing integer bit quantization on a quantized layer parameter so as to approximate a parameter matrix of the quantized layer. The present invention also provides a human face recognition device based on the residual-quantized convolutional neural network.

A cucumber disease identification method based on cucumber leaf symptom image processing comprises the steps that cucumber disease leaf scab images are segmented; cucumber disease leaf image identification features are extracted, dimensionality reduction is carried out on feature vectors, and at last cucumber disease identification is carried out. The method resolves the problems that the identification rate of cucumber diseases based on leaves is not high and identification effects are instable due to the facts that according to an existing cucumber disease identification method and technique, cucumber disease leaf image components are complex, scabs on disease cucumber leaves are irregular in arrangement and color, and shapes and colors of scabs of leaves of different diseases are not the same, and has the advantages of being high in feature extraction speed and identification rate, stable in identification effect, higher in practicability and the like.

The invention discloses a neural network acceleration method based on cross-resolution knowledge distillation. The method comprises the steps of acquiring high-resolution and low-resolution training samples; constructing high-resolution and low-resolution student networks; pre-training a teacher network through the high-resolution sample data; fixing teacher network parameters, and extracting teacher network output from the high-resolution image; extracting low-resolution image features by using a student network, and constraining the output features of the high-resolution teacher network andthe low-resolution student network to be consistent through cross-resolution distillation loss; in the test stage, extracting robust features from a low-resolution input image by using a student network. According to the invention, knowledge propagation between high-resolution and low-resolution fields is realized by using cross-resolution distillation loss; by extracting the feature accelerationnetwork from the low-resolution image, the calculation complexity is reduced, the discrimination capability and generalization capability of the depth rule are improved by using the prior knowledge ofthe high-resolution image, and the excellent recognition performance is maintained while the calculation complexity of the depth network is greatly reduced.

The invention discloses a hand-written character input method, which comprises: A, ,performing feature selection on prestored character classes and calculating a template in coarse classification; B, performing feature transformation on the prestored character classes and calculating a template in fine classification; C, acquiring a discrete coordinate sequence of a track point of an input character and adopting a smooth continuous function to adjust the size and shape of the hand-written input character and a centrobaric coordinate value; D, extracting feature to obtain a multidimensional feature vector of the hand-written character; E, selecting partial feature value of the hand-written input character, matching the hand-written input character with the template in coarse classification respectively, and selecting a plurality of candidate character classes with the maximum similarity; and F, performing feature transformation on the hand-written input character, matching the hand-written input character with the sample center of the candidate character classes selected from the template in fine classification, and determining the most similar character classes. The invention also discloses a hand-written character input system. The invention has higher speed of identifying the hand-written input character, and higher identification accuracy.

The invention discloses a polarization multi-feature based method for detecting an ocean floating radar target. The method comprises the following steps: (1) transmitting a pulse signal by a radar transmitter, receiving anchored data by a radar receiver, and constructing a receiving vector of a detection unit and a receiving vector of a reference unit in the anchored data under H0 hypothesis and H1 hypothesis; (2) respectively calculating two polarization features of the detection unit and two polarization features of the reference unit; (3) respectively calculating the relative Doppler peak height of the detection unit and the relative Doppler peak height of the reference unit; (4) respectively calculating the combined feature of the detection unit and the combined feature of the reference unit; (5) calculating the detection statistics of the detection unit; and (6) determining that a target exists in the detection unit and the H1 hypothesis is established if the detection statistics of the detection are more than zero, and determining a target does not exist in the detection unit and the H0 hypothesis is established if the detection statistics of the detection are not more than zero.

The embodiment of the invention provides an image processing method and device and electronic equipment, and the method comprises the steps: obtaining at least two to-be-processed images, wherein theat least two images are images corresponding to at least two sub-bands in a scanning synthetic aperture radar ScanSAR image; correcting the proportion of the at least two images by utilizing the geographic position information of the images to obtain at least two corrected images; performing feature extraction on the at least two corrected images based on an ORB method to obtain a feature extraction result; splicing the at least two corrected images according to the feature extraction result. Thus, the proportion of the images is corrected according to the geographic information of the images,the corrected images are spliced based on the ORB method, and rapid and high-precision splicing of the images is achieved.

The embodiment of the invention provides a malicious software detection method and device, equipment and a storage medium, and relates to the technical field of network information security, the method comprises the following steps: statically analyzing a binary file of to-be-detected software to obtain an assembly code and a function call graph of the to-be-detected software; converting the assembly code to obtain a semantic feature vector of each function in the to-be-detected software; combining the semantic feature vector with the function call graph to generate an attribute function call graph; and inputting the attribute function call graph into a graph neural network classification model to obtain malicious attribute information of the to-be-detected software. According to the malicious software detection method, the semantic features and the structural features of the binary program can be automatically extracted, the semantic features and the structural features are combined and judged through the graph neural network, the problems that in an existing detection method, feature representation is incomplete, and the missing report rate and the false alarm rate are high are solved, and the malicious software can be quickly and accurately detected.

A gesture recognition method based on a deep neural network and an attention mechanism belongs to the field of electronic information. The method comprises the following steps: firstly, introducing an ECA effective channel attention in a double-flow algorithm to enhance the attention of the double-flow algorithm on a gesture key frame, and respectively extracting space and time sequence characteristics in a dynamic gesture by utilizing a space convolutional network and a time convolutional network in the double-flow algorithm; secondly, a gesture frame with the highest attention degree is selected from the spatial stream through ECA, and corresponding hand posture features are extracted through the single-shot multi-frame detector technology; and finally, fusing the hand posture features with human body posture features and gesture time sequence features extracted from double flows, and then classifying and recognizing gestures. According to the method, verification is carried out on a Challarn2013 multi-mode gesture data set, the accuracy rate is 66.23%, and compared with a previous method that double-flow recognition is carried out on the data set only through RGB information, a better gesture recognition effect is obtained.

The invention relates to the field of central air conditioner energy consumption prediction, discloses a feature extraction method and system for energy consumption prediction, and aims to quickly screen input features of energy consumption prediction and improve the generalization performance of an energy consumption prediction algorithm. The method comprises the steps of collecting historical operation data of a to-be-analyzed central air conditioning system, and preprocessing the historical operation data to obtain an initial feature set; performing training according to the initial featureset to obtain a gradient boosting tree energy consumption prediction model, and calculating the contribution degree of each input feature; performing feature screening according to the contribution degree to obtain an optimized feature set; optimizing the gradient boosting tree energy consumption prediction model according to the optimization feature set, and obtaining a prediction value according to the optimized gradient boosting tree energy consumption prediction model; calculating a mean square error between the contribution degree and the predicted value; and screening the mean square error by adopting a preset feature screening termination condition to obtain an optimal feature set.

The invention discloses a finger-vein feature extraction method, a comparison method, a storage medium and a processor. The finger-vein feature extraction method includes such steps as convoluting a finger vein image with g (x, sigma 1) cos (k1x) in horizontal direction, convoluting finger vein image with g (y, sigma 1) cos (k2y) in vertical direction to obtain first intermediate result, convoluting of finger vein image with g (x, sigma 1) cos (k2y) in horizontal direction to obtain first intermediate result, convoluting the finger vein image with g (x, sigma 1) cos (k2y) in vertical directionto obtain first intermediate result, convoluting the the finger vein image with g (x, sigma 1) cos (k2y). The vein image is convoluted with g (x, sigma 1) sin (k 1 x) in horizontal direction and thenwith g (y, sigma 1) sin (k 2 y) in vertical direction to obtain the second intermediate result; Gaussian convolution of that vein image in the horizontal and vertical directions is perform and multiplied (shown in the description) to obtain a third intermediate result; The first intermediate result is subtracted from the second intermediate result, and the third intermediate sum is obtained to obtain the final result, that is, the filtering result of Gabor, that is, the Gabor feature of the image is extracted. The invention effectively reduces the calculation amount, improves the feature extraction speed, effectively reduces the calculation overhead and consumes the memory, meets the real-time requirement, and has low calculation complexity.

The invention discloses a hyperspectral image rapid filtering method based on a three-dimensional Gabor filter. The method includes the following steps: obtaining a three-dimensional hyperspectral image cube, and building a three-dimensional Gabor filter with different frequency amplitude, and different frequency methods and scale parameters after preprocessing; determining an optimal scale parameter of the three-dimensional Gabor filter through a cross verification method, then decomposing an imaginary part and a real part of the filter into a linear combination of one-dimensional low-order filter products, further obtaining characteristic cubes of the real part and the imaginary part, then determining modulus values, merging into a module value characteristic cube, connecting and mergingamplitude characteristic cubes with different frequency, amplitude and orientation parameters, obtaining characteristics for classification, and finally verifying classification accuracy.

The invention provides an image feature extraction method and device and electronic equipment, and relates to the technical field of image processing, and the method comprises the steps of obtaining a to-be-processed image and a preset convolution kernel, wherein the to-be-processed image comprises an original image or a feature map, and the data types of the to-be-processed image and the convolution kernel are both fixed-point types; performing convolution processing on the to-be-processed image based on the convolution kernel to obtain a convolution processing result used for representing image features of the to-be-processed image, wherein the data type of the convolution processing result is a fixed point type; involving floating point calculation and conversion between fixed points and floating points in the convolution processing process. The invention can effectively improve the feature extraction precision and speed.

The invention discloses a laser weapon light spot monitoring and tracking system, and the system comprises the steps: displaying a laser light spot, calculating the size of the light spot, and judging whether the laser light spot meets the irradiation requirements or not; when it is determined that the light spot meets the requirement, clicking an irradiation target, and tracking the position of the light spot; and evaluating the precision of the laser guided weapon according to the final position of the light spot. The technical problem of laser guided bomb precision evaluation in related technologies is solved, and the technical effects of being simple in structure, low in cost and high in portability are achieved.

The invention discloses a method for acquiring features of controlled drugs, which is applied to an electronic device. The method includes: acquiring original case data of an insurance institution database and a medical institution database; performing preprocessing on the original case data, and obtaining the preprocessed A data set; establish a classification model according to the forward characteristics, natural characteristics and backward characteristics of the controlled drug; extract the data of the controlled drug through the classification model; establish a specific controlled drug feature extraction according to the unique characteristics of the specific controlled drug model; and return the feature table of the specific controlled drug through the feature extraction model. The invention also provides an electronic device and a computer-readable storage medium. The present invention can quickly obtain feature sets of various controlled drugs by analyzing big data, and greatly improve the feature extraction speed and accuracy of different controlled drugs in different regions.

A cucumber disease recognition method based on cucumber leaf symptom image processing, first segment the cucumber disease leaf lesion image, then extract the image recognition features of cucumber disease leaf, then reduce the dimensionality of the feature vector, and finally identify the cucumber disease. The invention overcomes the existing methods and technologies for identifying cucumber diseases. Due to the complex image components of cucumber diseased leaves, the arrangement of diseased spots on cucumber diseased leaves is irregular, and the color is different, and the shapes and colors of leaf diseased spots of different disease types are not the same. The reason is that the recognition rate of cucumber diseases based on leaves is not high and the recognition effect is unstable. It has the advantages of fast feature extraction, high recognition rate, stable recognition effect and strong practicability.

The invention relates to a tunnel cross section feature extraction method and device, computer equipment, a storage medium and a computer program product, is applied to the technical field of tunnel three-dimensional point cloud data feature extraction, and is used for improving the accuracy of tunnel cross section feature extraction. The method comprises the following steps: sampling cross section data of a target tunnel cross section to obtain sampling data; inputting the sampling data into a parameter estimation model corresponding to the cross section of the target tunnel, and estimating parameters in a contour line fitting model corresponding to the cross section of the target tunnel through the parameter estimation model based on the sampling data to obtain estimated values of the parameters in the contour line fitting model; and inputting the estimated value into the contour line fitting model to obtain a contour line of the target tunnel cross section as a first target feature of the target tunnel cross section.

The invention discloses an image multi-scale feature extraction method based on a cellular neural network, and the method comprises the steps: firstly generating a plurality of pairs of feature maps in a binuclear recursion convolution mode based on a local binary constraint improved cellular neural network under the condition of neuron neighborhood L (r, p) sampling; compressing the state featuremap by using rotation invariant mapping and low-frequency mode combination; generating a single-scale joint mode proportional histogram of the image on the state feature map and the response featuremap according to a joint distribution mode statistical rule; performing softmax optimization on the joint histogram, and adding a standard variance component, so that an optimized single-scale featurevector can be obtained; and finally, connecting the plurality of single-scale vectors in series to obtain a multi-scale feature vector of the image.

In order to provide a face recognition method and device that can not only complete large-scale face recognition, but also reduce the amount of calculation to reduce hardware requirements and reduce the time required for training, the present invention provides a method based on residual quantized convolution The face recognition method of the neural network comprises the following steps: step S1, constructing a convolutional neural network model and training; step S2, preprocessing the target image and preprocessing the image to be determined; step S3, preprocessing the image to be determined The image and the preprocessed target image are sequentially input into the feature extraction model to obtain the feature vector to be determined and the target feature vector; step S4, to determine consistent face images according to the target feature vector and the vector to be determined, wherein step S1 includes setting the predetermined layer A step of approximating the parameter matrix of the quantization layer by performing integer bit quantization on the quantization layer parameters. The invention also provides a face recognition device based on residual quantization convolutional neural network.

The invention discloses a hand-written character input method, which comprises: A, ,performing feature selection on prestored character classes and calculating a template in coarse classification; B, performing feature transformation on the prestored character classes and calculating a template in fine classification; C, acquiring a discrete coordinate sequence of a track point of an input character and adopting a smooth continuous function to adjust the size and shape of the hand-written input character and a centrobaric coordinate value; D, extracting feature to obtain a multidimensional feature vector of the hand-written character; E, selecting partial feature value of the hand-written input character, matching the hand-written input character with the template in coarse classificationrespectively, and selecting a plurality of candidate character classes with the maximum similarity; and F, performing feature transformation on the hand-written input character, matching the hand-written input character with the sample center of the candidate character classes selected from the template in fine classification, and determining the most similar character classes. The invention alsodiscloses a hand-written character input system. The invention has higher speed of identifying the hand-written input character, and higher identification accuracy.

The invention discloses a VI waveform feature extraction method for user power consumption behavior perception analysis, and relates to the field of power grid operation and maintenance. At present, aiming at a power consumption behavior feature extraction technology at home and abroad, starting from the perspective of load features, universality is poor, harmonic waves are extracted to serve as feature data, and variable speed drive type loads cannot be further distinguished. A current and voltage trajectory curve is collected; the closed area, the average curve distortion degree and the curve self-intersection point number of a trajectory curve are considered; five characteristics of a current and voltage trajectory curve, the slope of a near zero point of the middle section of an average curve and the clockwise or anticlockwise average curvature of a VI trajectory are considered, and mathematical mode feature extraction is carried out, thus obtaining a feature data set of user behavior perception analysis. The invention is high in feature extraction speed and easy to implement, multiple features jointly participate in user behavior perception, and the accuracy is high. The method can provide data support for user load identification, and has good economic benefit and practical value.