37results about How to "Improve recognition time" patented technology

The invention discloses a three-dimensional convolutional neural network (3D CNN) based video classifying method and belongs to the technical field of video processing. According to the method, a video is sampled at equal intervals to obtain a plurality of video segments, a video database is amplified, three-dimensional video segments are directly input into a 3D CNN, and time domain and space domain characteristics of the video are extracted, so that the limitation of a conventional video classifying method in manually selecting video characteristics and video modeling modes is improved. A parallel distributed 3D CNN multi-classification model lowers the complexity in learning the 3D CNN and enables a classification system to realize distributed parallel computation more conveniently. Relatively high identification rate can be achieved with only fewer video segments based on a 3D CNN multi-classification system, and videos not belonging to any type can be classified into new type, so that the classification error of the new type is avoided.

The invention relates to the sensor data processing technology field and particularly relates to a data fusion and intelligent searching processing method for multiple data sources, and an application. According to the method, sensors are arranged in a planar layout, the sensors are arranged on one same plane to form a sensing network, data fusion comprises data fusion through multiple same-type sensors and data fusion through multiple different-type sensors, data characteristics and data types of the sensors are acquired by employing a polling mode, redundancy processing on the data is carried out when acquisition is carried out by an acquisition node, and an self-adaptive algorithm based on batched estimation is employed. Through the method, the data of each sensor is dynamically acquired, the identification time of a system for the sensor data is prolonged, data precision is improved, data accuracy is improved, identification reliability of the system is improved, the method is suitable for occasions having a relatively large quantity of same-type sensors, relatively large data amount and same-type-sensor-in-one-same-plane layout, and data fusion in plantation environments and farmland climate monitoring multi-sensor environments can be realized through the method.

The invention provides a human face emotion analysis method and system based on multi-task learning and deep learning. The method comprises the steps of learning preset analysis task convolution layers in a human face database by using a convolutional neural network to obtain a human face analysis model; acquiring a to-be-analyzed human face image, analyzing the to-be-analyzed human face image by using a human face detection algorithm and extracting human face areas in the to-be-analyzed human face image; performing prediction on the to-be-analyzed human face image by using the human face analysis model and obtaining emotion information corresponding to each human face area in the to-be-analyzed human face image. According to the invention, the concept of multi-task learning is applied to a convolutional neural network and multiple kinds of analysis tasks related to human faces can be identified by using the same analysis model, so that the size of an analysis model is reduced and identification is accelerated; different parts of a human face are described by using different convolution layers, so that each convolution layer focuses on one single task and a better identification effect can be achieved.

A shift control ECU includes the steps of storing a position of a shift lever; starting a count-up timer when the shift lever comes to a neutral position; and recognizing a driver's request for the neutral position when a count-up timer value attains TN(1) if the stored shift lever position is a home position and when a count-up timer value attains TN(2) if the stored shift lever position is not the home position. Here, TN(1) is set to be smaller than TN(2).

The invention discloses a three-dimensional palm print identification method based on a partitioning statistical characteristic and joint expression. A sample image of a three-dimensional palm print, is equally divided into a plurality of subareas, for each subarea, the surface type of the palm print is calculated firstly, then the surface types of palm prints in each subarea is counted by use of a histogram, finally, surface type histograms of each subarea are spliced to be taken as a characteristic description operator of a palm depth image, and classification is carried out by use of a joint expression frame, such that the identification efficiency and the accuracy are improved, the method can be applied to occasion with strict requirements for identity identification, and the problem of aligning deviations among multiple three-dimensional palm print samples is effectively solved.

The invention provides a rapid, high-robustness and autonomous fixed star identification method, which comprises the following steps: firstly, finding a navigation star pair with a highest matching frequency from all observed star pairs to generate a matching matrix; secondly, confirming the matching relation between every two matching pairs in the matching matrix to generate a matching confirmation matrix; finally, calculating a maximum matching group according to the matching confirmation matrix to obtain an identification result. The method of the invention has the advantages of high identification speed, good robustness, high identification probability, low mis-identification rate, low algorithm realization cost, and high identification error tolerance rate.

The invention discloses a quick near-infrared brain-computer interface method. The method is mainly characterized in that a multi-channel near-infrared optical brain signal collecting system is used for obtaining multi-channel near-infrared optical brain signals generated by a user in the brain-computer interface task executing process in real time, after the signals are pre-processed, feature space transformation is carried out to extract feature vectors, and mode recognition is carried out to judge the type of a brain-computer interface task executed by the user. The method has the advantages that the recognition rate and recognition speed of brain-computer interface tasks are obviously improved compared with a traditional near-infrared brain-computer interface method which carries out recognition with amplitude as feature vectors.

The invention discloses a method and a system for quickly recognizing feature combinations in high-dimensional data. The method and the system have the advantages that cross validation and classification errors which are measured by a leave-one-out process can be minimized while the number of selected features is minimized, and the feature combinations in the high-dimensional data can be ultimately quickly recognized by means of modeling, so that the method can be used for statistically analyzing the high-dimensional data, and the method and the system have wide application prospects in fields of data mining, machine learning, artificial intelligence, biomedicine and the like.

A shift control ECU includes the steps of storing a position of a shift lever; starting a count-up timer when the shift lever comes to a neutral position; and recognizing a driver's request for the neutral position when a count-up timer value attains TN(1) if the stored shift lever position is a home position and when a count-up timer value attains TN(2) if the stored shift lever position is not the home position. Here, TN(1) is set to be smaller than TN(2).

The invention provides a robot video call control method and device, and a terminal and mainly relates to a smart robot capable of realizing a video call through remote control. The method comprises the steps of establishing the video call with a call party and transmitting video stream obtained by a local to the call party; receiving a target-seeking instruction initiated by the call party, analyzing target object information contained in the target-seeking instruction, and determining corresponding target feature information according to the target object information; when a target object is not captured, starting a walking device to carry out local movement, carrying out image identification on the video stream in a movement process, and determining an image containing the target feature information, thereby capturing the target object; and after the target object is captured, controlling the walking device to enable the local to keep a preset distance range with the target object. According to the method, the device and the terminal, the walking device fits with a camera unit, the target object is rapidly captured in the movement process through an image identification technology, the video call is carried out, and the aloneness of a child is reduced through a man-machine interaction function.

The invention relates to a new geometric invariant (namely feature number)-based shape descriptor, and belongs to the field of computer vision. The shape descriptor is an outline-based shape descriptor which combines a global feature and a local feature. A geometric graphic feature point-based method for establishing the shape descriptor comprises the following steps of: sampling convex hulls of an image uniformly to obtain a series of sample points; and selecting three points sequentially among the sample points according to the definition of feature numbers to obtain a series of feather numerical values, so that a feature vector, namely the shape descriptor of the image is formed. According to the shape descriptor, the information of an internal outline of a shape is utilized fully; the content which is described by each component of the descriptor can cover a certain area of the shape, so that the capacity and accuracy degree of describing the shape are improved; the descriptor is used as a projection invariant, so that the shape descriptor can be suitable for various geometric transformations; and for the interference of noise of different degrees, the shape descriptor is high in stability. In the matching stage, a point-to-point dynamic matching mode is replaced by a first-point matching mode, so that recognition time is prolonged greatly. The feature number-based shape descriptor has high compactness and stability and high practical value.

The invention discloses a paper money identification system, which comprises data acquisition units, data storage units and data identification units. The number of the data storage units is N and the number of the data identification units are more than or equal to N minus 1, wherein N is a natural number more than or equal to 2. The data storage units and the data identification units are connected with the data acquisition units respectively. The data identification units and the data storage units are connected. The data acquisition units, the data storage units and the data identification units acquire, storage and identify data information of different paper moneys respectively, so that the system can carries out different operations to a plurality of pieces of paper moneys at the same time and allows the different operations to be coincide, different from the identification method that identifies one piece of paper money each time. On the premise of ensuring the paper money conveying speed, the system greatly improving time utilization rate during paper money identification and contributes to the improvement of the precision of paper money identification. In addition, the invention also discloses a paper money identification method using the system.

The invention discloses a single-sample face recognition method based on Gabor feature extraction and spatial transformation, and mainly solves the problem that a traditional face recognition method can not be applied since a within class scatter matrix is zero under a condition that only single training sample image is in the presence. The method comprises the following steps that: adopting a Gabor wavelet to extract a spatial feature vector from an original single sample image; then, carrying out fusion on the extracted spatial feature vector with an original spectral feature vector, utilizing a feature spatial transformation method to carry out low-dimension feature space transformation on a fused feature matrix, and transforming the fused feature matrix to a low-dimension subspace; and finally, utilizing a nearest neighbor classifier to finish recognition. By use of the method, single-sample face recognition can be accurately finished, recognition accuracy is improved, and calculation cost is lowered. Compared with the prior art, the face recognition method which is put forward by the invention exhibits better effectiveness and robustness.

The invention discloses a three-dimensional ear recognition method based on block statistic features and dictionary learning sparse representation classification. A three-dimensional ear sampling image is equally divided into a plurality of sub regions; as for each sub region, an ear surface type is firstly calculated, a histogram is then used for counting the ear surface type in each sub region, surface type histograms of all sub regions are finally spliced to serve as a feature description operator for the ear depth image, a dictionary learning sparse representation framework is used for classification, and thus the recognition efficiency and the accuracy are improved. The method of the invention can be used in situations with strict requirements on identity recognition, and can effectively solve the alignment deviation problem among multiple three-dimensional ear sampling.

The invention discloses a star sensor artificial star screening method based on angular distance screening. The method comprises the following steps: (1) all star points in a first frame of a star mapare recorded as initial star points; (2) starting from a second frame, corresponding relations between each star point and the initial star points are established by utilizing a star tracking method;(3) the angular distances between other frames and the initial frame are compared, the angular distances are screened according to angular distance changes, and an angular distance screening matrix is updated; (4) star pairs corresponding to the screened angular distances are used to vote for each star point, and the star points with the excessively low vote numbers are screened; and (5) each frame is output by the screened star points and is sent to a subsequent star map recognition module. The method has the following advantages: (1) the convergence success rate is high, and the convergencesuccess rate is over 98% within 600 artificial stars; (2) the convergence speed is high, and all the artificial stars can be screened by only at most 10 frames of the star map on average; (3) the adaptability to the number of the artificial satellites is high, and all the artificial stars can be screened from the star map with no less than 700 artificial stars; and (4) the screening of the artificial stars does not depend on shape information of the star points.

A shift control ECU includes the steps of storing a position of a shift lever; starting a count-up timer when the shift lever comes to a neutral position; and recognizing a driver's request for the neutral position when a count-up timer value attains TN(1) if the stored shift lever position is a home position and when a count-up timer value attains TN(2) if the stored shift lever position is not the home position. Here, TN(1) is set to be smaller than TN(2).

The invention relates to the technical field of data updating, and provides a training set updating method and device for a detection model, and the method comprises the steps: obtaining a sample picture, identifying a field region from the sample picture according to a field detection algorithm, and labeling a first detection box; expanding the first detection frame towards the periphery according to a set proportion to obtain a second detection frame, and cutting the sample picture according to the second detection frame to obtain a target field picture; using a crowdsourcing platform to identify whether the second detection box of the target field picture is misplaced or not, and carrying out misplacement marking; receiving an identified target field picture returned by the crowdsourcing platform, and if a labeling user of the crowdsourcing platform corrects the misplaced target field picture according to the misplaced label, determining that the field detection algorithm has an error; and updating the corrected target field picture to a training set of a detection model. The detection capability of the detection model can be improved.

The face emotion analysis method and system based on multi-task learning and deep learning provided by the present invention include using a convolutional neural network to learn the convolution layer of a preset analysis task in a face database to obtain a face analysis model; A face image, using a face detection algorithm to analyze the face image to be analyzed, extracting the face area in the face image to be analyzed; using the face analysis model to predict the face image to be analyzed, and obtaining Emotional information corresponding to each face area in the face image to be analyzed. The present invention applies the concept of multi-task learning to the convolutional neural network, so that a variety of face-related analysis tasks can be identified with the same analysis model, which can reduce the size of the analysis model and speed up the recognition time. In addition, the present invention uses different convolutional layers to describe different parts of the face, so that the tasks of each convolutional layer are relatively single and specialized, and better recognition effects can be achieved.

The invention discloses a video classification method based on a three-dimensional convolutional neural network, belonging to the technical field of video processing. This method samples the video into multiple video segments at equal intervals, expands the video database, directly inputs the 3D video segment into the 3D CNN, and extracts the temporal and spatial features of the video at the same time, which improves the traditional video classification method due to manual selection of video. Limitations of feature and video modeling methods. The parallel distributed 3D CNN multi-classification model reduces the complexity of 3D CNN learning, and at the same time makes it easier for the classification system to implement distributed parallel computing. The 3D CNN-based multi-classification system can achieve a high recognition rate with only a small number of video segments, and can classify videos that do not belong to any category into new categories, avoiding classification errors for new categories.

The invention discloses a single-sample face recognition method based on Gabor feature extraction and space transformation, which mainly solves the inability of the traditional face recognition method due to the fact that the intra-class scatter matrix is ​​zero under the condition of only a single training sample image. application problem. This method uses Gabor wavelet to extract the spatial feature vector from the original single-sample image, then fuses the extracted spatial feature vector and the original spectral feature vector, uses the feature space transformation method to perform low-dimensional feature space transformation on the fusion feature matrix, and converts it to Transform into a low-dimensional subspace, and finally, use the nearest neighbor classifier to complete the identification. The method of the invention can accurately complete the recognition of a single-sample human face, improves the recognition precision and reduces the calculation cost. Compared with the prior art, the face recognition method proposed by the present invention is more effective and robust.

The invention discloses a three-dimensional human ear recognition method based on block statistical features and dictionary learning sparse representation classification, which divides the sampled image of the three-dimensional human ear into several sub-regions, and first calculates the surface of the human ear for each sub-region type, and then use the histogram to count the surface type of the human ear in each sub-region, and finally stitch the surface type histograms of each sub-region together as the feature description operator of the depth image of the human ear, and use the dictionary to learn the sparse representation framework for classification , so as to improve the recognition efficiency and accuracy. The invention can be used in occasions with strict requirements on identity recognition, and effectively solves the problem of alignment deviation among multiple three-dimensional human ear samples.

The invention relates to a spectrum fusion intelligent target detection method which comprises a first step and a second step which are sequentially carried out. The first step is to carry out infrared image target recognition on a target to be recognized so as to obtain a first recognition result, if the first recognition result can reach an expected recognition rate, the operation is finished, if the first recognition result cannot reach the expected recognition rate, entering the second step; the second step is that the target to be recognized is recognized by the infrared spectrum associated target to obtain a second recognition result, the second recognition result is taken as a final recognition result, and the operation is finished at the same time. The spectrum fusion intelligent target detection method and light path switching device thereof not only has the recognition result and the recognition cost, but also has high cost performance and simple operation.

The invention discloses a method and a system for quickly recognizing feature combinations in high-dimensional data. The method and the system have the advantages that cross validation and classification errors which are measured by a leave-one-out process can be minimized while the number of selected features is minimized, and the feature combinations in the high-dimensional data can be ultimately quickly recognized by means of modeling, so that the method can be used for statistically analyzing the high-dimensional data, and the method and the system have wide application prospects in fields of data mining, machine learning, artificial intelligence, biomedicine and the like.

The invention relates to the sensor data processing technology field and particularly relates to a data fusion and intelligent searching processing method for multiple data sources, and an application. According to the method, sensors are arranged in a planar layout, the sensors are arranged on one same plane to form a sensing network, data fusion comprises data fusion through multiple same-type sensors and data fusion through multiple different-type sensors, data characteristics and data types of the sensors are acquired by employing a polling mode, redundancy processing on the data is carried out when acquisition is carried out by an acquisition node, and an self-adaptive algorithm based on batched estimation is employed. Through the method, the data of each sensor is dynamically acquired, the identification time of a system for the sensor data is prolonged, data precision is improved, data accuracy is improved, identification reliability of the system is improved, the method is suitable for occasions having a relatively large quantity of same-type sensors, relatively large data amount and same-type-sensor-in-one-same-plane layout, and data fusion in plantation environments and farmland climate monitoring multi-sensor environments can be realized through the method.

The present invention relates to the laser technology field, in particular to a coherent domain-based chaotic laser external cavity time delay identification method which solves the technical problemsthat a conventional chaotic laser external cavity time delay identification method mainly carries out the two-dimensional space mathematical analysis on the intensity information of a chaotic signalin the time domain or the frequency domain, when facing various technical methods of hiding or suppressing the time delay information provided at present, often cannot identify the time delay information effectively. The chaotic laser external cavity time delay identification method provided by the present invention carries out in an optical coherent domain and avoids a photoelectric conversion process of converting the chaotic laser into an electric signal in a time domain and frequency domain analysis method, and the whole chaotic laser external cavity time delay identification process is carried out in the optical domain, so that the chaotic laser external cavity time delay identification time is improved greatly. The method provided by the present invention is more accurate to identifythe chaotic laser external cavity time delay information, and can identify the chaotic laser external cavity time delay that cannot be identified effectively by the conventional method.

The invention discloses a quick near-infrared brain-computer interface method. The method is mainly characterized in that a multi-channel near-infrared optical brain signal collecting system is used for obtaining multi-channel near-infrared optical brain signals generated by a user in the brain-computer interface task executing process in real time, after the signals are pre-processed, feature space transformation is carried out to extract feature vectors, and mode recognition is carried out to judge the type of a brain-computer interface task executed by the user. The method has the advantages that the recognition rate and recognition speed of brain-computer interface tasks are obviously improved compared with a traditional near-infrared brain-computer interface method which carries out recognition with amplitude as feature vectors.

The invention discloses a three-dimensional palm print identification method based on a partitioning statistical characteristic and joint expression. A sample image of a three-dimensional palm print, is equally divided into a plurality of subareas, for each subarea, the surface type of the palm print is calculated firstly, then the surface types of palm prints in each subarea is counted by use of a histogram, finally, surface type histograms of each subarea are spliced to be taken as a characteristic description operator of a palm depth image, and classification is carried out by use of a joint expression frame, such that the identification efficiency and the accuracy are improved, the method can be applied to occasion with strict requirements for identity identification, and the problem of aligning deviations among multiple three-dimensional palm print samples is effectively solved.

The invention discloses a tree search RFID anti-collision method capable of reducing communication complexity, and solves the problems of large data communication traffic, identification time delay and the like in an RFID search tree anti-collision algorithm. The method comprises the following steps: introducing a prefix length register and a response mark register into a label, storing prefix number information in a reader stack area, enabling the reader to send the prefix length information to classify and search the labels, wherein the reader and the labels do not send the serial numbers identified by each other any more, so that the communication data volume is effectively reduced. Simulation results show that compared with a traditional binary tree search anti-collision algorithm, thealgorithm can obviously reduce the system communication complexity and improve the tag search rate.

The invention relates to the fields of intelligent electricity utilization and energy efficiency monitoring, and more particularly to a method for identifying residential load electricity utilization based on current signal separation. The method includes the steps of firstly, storing current signal data when appliances are running independently; recording existing signals in a circuit as one way of source signals I0 and storing current signal data thereof; after detecting the start of a next appliance, acquiring mixed signals and using a two-step iterative contraction threshold algorithm to recover two-way separation signal frequency domain values therefrom; and calculating a similarity coefficient between the separation signals and the frequency domain values of the stored signals, separating only two ways of signals from the acquired signals, one way of which is independent current signals of a new running load, and the other way of which is the mixed current of the previous moment, and recording the mixed signals at the moment as one way of source signals I0. According to the invention, operation conditions of each single load can be effectively separated from the mixed data when a plurality of loads run simultaneously under a non-invasive electricity data acquisition condition, and the method is an effective realization method of non-invasive load monitoring technology.

The invention provides a rapid, high-robustness and autonomous fixed star identification method, which comprises the following steps: firstly, finding a navigation star pair with a highest matching frequency from all observed star pairs to generate a matching matrix; secondly, confirming the matching relation between every two matching pairs in the matching matrix to generate a matching confirmation matrix; finally, calculating a maximum matching group according to the matching confirmation matrix to obtain an identification result. The method of the invention has the advantages of high identification speed, good robustness, high identification probability, low mis-identification rate, low algorithm realization cost, and high identification error tolerance rate.