516 results about "Sample selection" patented technology

The invention relates to the technical field of image information processing, in particular to a multiclass image classification method based on active learning and semi-supervised learning. The method comprises five steps: initial sample selection and classifier model training, BvSB active learning sample selection, CST semi-supervised learning, training sample set and classifier model updating and assorting process iteration. Through the operations of BvSB active learning sample selection, CST semi-supervised learning and SVM classification, the invention has efficient image classification effect under the condition of less manual tagging, does not increase overmuch computation burden, can quickly provide classification effects and also can take consideration of the demand of a classification system on the computation complexity.

A repeatable read-out (RRO) detector employs one or more digital interpolators to interpolate asynchronous sample values that represent RRO data. The asynchronous sample values are read from a recording medium and generated by an A/D converter at a symbol rate, and the interpolators generate interpolated samples at at least one time in between the asynchronous sample value times. Each interpolated sample corresponding to some phase relative to that of the sample values generated by the A/D converter. The RRO detector receives 1) the asynchronous samples at symbol rate and 2) the interpolated samples to efficiently detect the encoded RRO data. An RRO address mark indicates when detection of encoded RRO data starts, and is employed to select those samples suitable for RRO data detection. Detection of the RRO address mark employs peak detection among filtered asynchronous and interpolated samples. The process of peak detection adjusts the current best phase for sample selection. When the RRO address mark is found, the corresponding best phase corresponds to either asynchronous sampled values or interpolated samples that are subsequently selected for RRO data detection, termed best samples. Once the best phase is selected, the RRO data detector uses that information along with RRO encoding constraints to decode the encoded RRO data from the best samples.

An apparatus for and method of performing orthogonal light sheet microscopy (OLM) and computer optical tomography (COT) simultaneously in a single device are provided. The dual-mode imaging microscope allows for the use of both OLM and COT in a single instrument. This dual-mode device will allow researchers to have access to both types of microscopy, allowing access to the widest possible selection of samples, and improved imaging results. In addition, the device will reduce the high costs and space requirements associated with owning two different imagers (OLM and COT).

The invention discloses a sample-training-based non-stationary clutter suppression method of a vehicle-mounted radar, which relates to vehicle-mounted radar technology. The method comprises the steps of estimation of a clutter covariance matrix based on combined time-dimension sample training strategies, application of a self-adaptive weight, and coherence stack to output signals. The method specifically comprises the following steps of: inputting raw echo data; compressing and windowing pulses in a distance dimension; segmenting slow-time-dimension data; selecting quick-slow time dimension training samples; estimating the clutter covariance matrix; calculating and applying the self-adaptive weight; and carrying out coherence stack on the output signals. According to the method, the sample training strategies are changed under an STAP (space-time adaptive processing) time domain dimension reducing structure in light of the clutter range dependence of the vehicle-mounted radar, thus the estimation precision of the clutter covariance matrix can be effectively improved, and the clutter suppression performance of a main lobe is improved as well. The sample-training-based non-stationary clutter suppression method shows high robustness in engineering application, and is particularly applicable to detection on a slow moving object.

A method for managing sample selection bias is disclosed. Embodiments of the method can include automatically determining an unbiased estimate of a distribution from occurrence data via a special purpose processor. The occurrence data that is utilized in determining the estimate may have an occurrence data sample selection bias that is substantially equivalent to a background data sample selection bias associated with background data. Additionally, the occurrence data may be related to the background data, and the background data may be chosen with the background data sample selection bias. Furthermore, the occurrence data may represent a physically-measurable variable associated with one or more physical and tangible objects or substances.

The invention discloses a remote sensing image classification method based on image block active learning and belongs to the technical field of image information processing. The method comprises the following steps of remote sensing image blocking, initial sample selection, classifier model training, active learning sample selection, training sample set and classifier model updating, classification process iteration; image block classification prediction, and conversion of a block classification result into a pixel classification result. The remote sensing image classification method serves an image block as a research object, compared with a traditional remote sensing image classification method based on active learning of pixel points, under the same experiment condition, the classification result of the image is more accurate, a block sample screened out by the active learning can more rapidly and accurately conduct manual annotating, constitutive properties of the classified image are stronger, spots directly brought by classification of the pixel points are greatly reduced, and the better visual effect is brought to people.

A sampling strategy using genetic algorithms (GA) in engineering design optimization is disclosed. A product is to design and optimize with a set of design variables, objectives and constraints. A suitable number of design of experiments (DOE) samples is then identified such that each point represents a particular or unique combination of design variables. The sample selection strategy is based on genetic algorithms. Computer-aided engineering (CAE) analysis or analyses (e.g., finite element analysis, finite difference analysis, mesh-free analysis, etc.) is/are performed for each of the samples during the GA based sample selection procedure. A meta-model is created to approximate the CAE analysis results at all of the DOE samples. Once the meta-model is satisfactory (e.g., accuracy within a tolerance), an optimized “best” design can be found by using the meta-model as function evaluator for the optimization method. Finally, a CAE analysis is performed to verify the optimized “best” design.

The invention provides a black box antagonistic attack defense method based on sample selection and model evolution. The method includes the following steps: 1) using a sample selector to randomly select partial samples from the multiple types of samples to be input into various attack models to generate a large number of counter samples; 2) calculating the attack effect of the counter samples, and analyzing the attack effect of different input samples and the attack model; (3) updating the number of different samples selected by different samples in the attack model and the sample selector according to the attack effect to make the newly generated counter sample have a better attack effect, meanwhile, updating the counter sample pool, storing the several counter samples with the best attack effect, and outputting the counter samples with the best attack effect in the pool to serve as the final result of the current evolution; and 4) training a large number of training output results and normal samples, so that the attack can be defended. According to the invention, the defense capability of the black box model can be improved.

The invention provides a classifier training method, a device, an electronic device and a computer-readable medium, belonging to the technical field of the Internet. A method for train that classifierinclude training labeled samples in a plurality of data feature to obtain a primary classifier; Inputting an unlabeled sample from the plurality of data features to the primary classifier for training to output a label of the unlabeled sample, and calculating a probability that the unlabeled sample labels the label; According to the probability of the label, a part of the unlabeled samples is selected for expert labeling, and the samples labeled by the expert are retrained to obtain a new classifier. This method is based on active learning to update the classifier. Combined with a large number of data and artificial experience, the classifier model can make the classification results more accurate and improve the classification effect. There are only a small number of labeled samples anda small number of manual labeling in the training samples, which can quickly process the classification of a large number of unlabeled samples and improve the efficiency.

The invention discloses a target detection method based on full-automatic learning. The method comprises the following steps: carrying out the training of a model through employing a small-scale manual labeling data set after preprocessing and a deep neural network, carrying out the fine adjustment of the model trained through an Imagenet data set, and obtaining a deep model; carrying out reasoning prediction on the pseudo-labeled part of the original large-scale image data set by utilizing the deep model, removing repeated prediction of the same target after carrying out non-maximum suppression, and respectively storing bounding boxes and confidence of prediction results according to categories. Through self-supervised pseudo labeling and active learning sample selection, the informationentropy and the divergence degree predicted by the deep neural network are learned in a combined manner, unlabeled samples are sorted according to the weight, and pseudo labels are allocated to high-confidence samples ranked at the top. The objective of the invention is to solve the problems of too high labor cost of common target detection and labeling and poor mobility and adaptability of a training model in an actual scene.

The invention provides an active learning based network data anomaly detection method, relating to an SVM (support vector machine) based detection method. The method solves the problems of more training samples, high complexity and poor classification feature of SVM in the traditional network data anomaly detection method. The method comprises the following steps: 1, choosing C points from unmarked samples as candidate cluster centers, carrying out iteration optimized clustering operation on candidate unmarked sample set A, and choosing representative samples from iteration clustering results to construct a training sample set B; 2, training an SVM on the training sample set B to obtain a training hyperplane; 3, choosing a sample which can best improve classification feature from the candidate unmarked sample set A according to sample choosing standards, making the sample with classification and then adding to the training sample set B; 4, retraining the SVM on the updated training sample set B; and 5, ending if detection precision reaches a preset value; if not, returning the step 3. The algorithm can effectively lower complexity if applied to anomaly detection research.

A computer vision dating system analyzes combinations of face features of the system's user's photographs and recommends potential dating partners. A user selects preferred and not-preferred faces from a sample of other user's pictures. The system analyzes the features of the preferred and not-preferred faces comparing the combinations of features in both categories with the features of other users in the database to find the users that most match the collective features preferred by the user. These pictures are presented to the user. Data from the user's profile input are analyzed to automatically generate the sample pictures from which the user selects his/her preferences. As the users are presented pictures after their sample selection, they can continue to select and reject pictures allowing the system to learn and refine the combinations of features and better locate those that most conform to a user's most preferred photo images.

An improved data recovery circuit based on an oversampling technique wherein intersymbol interference (ISI) is compensated. A detection circuit is connected at the output of a conventional recovery circuit. The recovered data is applied to the detection circuit which includes flip-flops to memorize the previous state of the recovered data when no data transition is detected within a predefined number of clock periods. The detection circuit detects sequences of a predetermined number of consecutive identical bits which indicates the presence of ISI. It generates a feedback signal that is applied to the decision circuit and to the data sample selection circuit to shift the selection of a data sample of one position to compensate ISI.

An improved data recovery circuit based on an oversampling technique to select the best data sample to be kept as the data to recover that is only based on accumulating the data edges (or transitions). The incoming serial data stream with jitter is oversampled in an oversampling circuit by means of the multiple phases of a reference clock (clk) to produce data samples. Each sample is compared to the sample(s) collected with the next clock phase(s) in an edge detector circuit to determine the presence of a data edge and the edge information is stored and accumulated in a data edge memory. A selection determination circuit uses the memorized edge information to indicate which sample is the farthest from the data edges. A selection validation circuit validates the selection to avoid false determination due to jitter and skew.

Layer processing to grow a layer structure upon a substrate surface comprises supplying a vapor mixture stream to the substrate (28) to deposit constituents, monitoring growth with an ellipsometer (12) and using its output in real-time growth control of successive pseudo-layers. A Bayesian algorithm is used to predict a probability density function for pseudo-layer growth parameters from initial surface composition, growth conditions and associated growth probabilities therewith, the function comprising discrete samples. Weights are assigned to the samples representing occurrence likelihoods based on most recent sensor output. A subset of the samples is chosen with selection likelihood weighted in favor of samples with greater weights. The subset provides a subsequent predicted probability density function and associated pseudo-layer growth parameters for growth control, and becomes a predicted probability density function for a further iteration of pseudo-layer growth.

The invention discloses a method and a system for change detection based on a multi-temporal remote sensing image and GIS data. The method is aimed at solving problems existing in segmentation, classification and change detection methods based on multi-temporal images and GIS data. The method includes the following steps: step 1, segmentation of a remote sensing image and extraction of pixel-levelchange information, mainly including remote sensing image segmentation and pixel-level change detection; step 2, object-oriented remote sensing image classification, mainly including change polygon decision, automatic sample selection and object-oriented SVM decision tree classification model; step 3, change detection after the object-oriented classification. Compared with existing methods of extracting change information using remote sensing images and GIS data, the method can realize automatic and large-area application change information detection method flow, can fully excavate relationship between remote sensing images and GIS data, utilizes attribute information such as type of the GIS data, and improves accuracy of change information extraction.

The invention discloses a transfer learning design method and system based on domain adaptation under a multi-example multi-label framework. According to the invention, multi-example multi-label learning and transfer learning are unified into one framework, source domain data samples and target domain data samples are effectively utilized for correlation statistics, and the source domain samples can be effectively used in the learning of target domain tasks; the characteristics of a source domain data sample set and a target domain data set in RKHS are utilized, a two-step domain adaptation process formed by sample weighting and a sample selection mechanism based on clustering is utilized, so that the learning of target tasks has enough training samples weighted and selected from the source domain set; and a miFV algorithm is utilized to convert multi-examples into a single example, and the calculation cost problem of domain adaptation is solved.

The invention provides an on-line detection method for particle size and shape of stacked aggregate. The method comprises the following steps: directly performing image acquisition on the stacked aggregate under an actual production state; processing an acquired stacked aggregate image; performing geometric characteristic analysis on the processed stacked aggregate image, and calculating a geometric characteristic of each aggregate particle in the stacked aggregate image; according to the geometric characteristic of each aggregate particle in the stacked aggregate image, analyzing to obtain particle size statistics information and particle shape distribution information of the stacked aggregate. According to the on-line detection method for the particle size and shape of the stacked aggregate, sample selection detection of the aggregate is avoided, and the simultaneous aggregate particle size and shape on-line detection can be realized under the actual production state, so that aggregate particle size and shape information in the actual production can be effectively, accurately and timely provided on line.

A receiver architecture for processing spread spectrum signals. The receiver has an RF front end to receive and down convert a broadcast signal to an intermediate frequency carrier. The IF signal is digitized and provided to a processor (which may be a software-driven DSP, an ASIC or other embodiment) for processing. A given IF carrier is removed and the signal is low pass filtered. The signal is provided to a number of channels, each, for example, correspond to a unique transmitter. On each channel the sample rate is reduced to a predetermined fixed rate with timing mismatch compensated. The Doppler frequency shift, as estimated for the channel, is removed succeedingly. A locally generated copy of the spreading code used by the transmitter is applied to the carrier and Doppler removed signal at the predetermined fixed sample rate. The de-spread signal is used to provide estimates of the Doppler shift and for subsequent sample selection. Pseudo-range and delta pseudo-range estimates from each channel are used to estimate, for example, the receiver's position.

The invention provides a tumble detection method and system. The method comprises an off-line training phase and on-line detection phase. In the off-line training phase, a plurality of sub-classifiers are trained combined with random sample selection and random classification attribute. In the on-line tumble detection phase, a final determination result is given based on the classification results of the plurality of sub-classifiers. Experiments show that through the provided tumble detection method, an accuracy rate of 95.2%, sensitivity of 90.6% and specificity of 93.5% can be obtained, the provided tumble detection method is better than a tumble detection method based on SVM and BP neural networks obviously, tumble behaviors are detected accurately, and the tumble detection method has strong generalization capability and robustness.

The invention relates to an intelligent extraction method of a build-up area on the basis of nighttime light data. The intelligent extraction method comprises the following steps: adopting an adaptive particle swarm optimization algorithm to realize the optimal selection of the selection parameters of a VIIRS (Visible Infrared Imaging Radiometer Suite) nighttime light image sample and a MODIS (Moderate Resolution Imaging Spectroradiometer) normalized difference vegetation index image sample; adopting a region growing algorithm based on SVM (Support Vector Machine) classification to finish SVM model training, and adopting a cross validation method to carry out precision validation on the model; and according to an optimized parameter, determining a city sample and a non-city sample, and adopting the region growing algorithm based on the SVM to extract a city built-area range. By use of the intelligent extraction method, from a sample selection source, the adaptive optimization of a sample selection parameter is carried out, and the SVM and the region growing algorithm are adopted to improve processing efficiency and precision for the nighttime light data to extract the nighttime light data.

The invention discloses a digital beam forming method of a broad band, which includes the steps of time division multiplexing quadrature sampling with intermediate frequency, quadrature demodulation, low pass filtering, interpolation delay, phase rotation and beam summation, etc. The beam forming method of the invention puts emphasis on that a single sample circuit is used for carrying out the time division multiplexing quadrature sampling with intermediate frequency on multipath bandpass signals; an FIR wave filter is used for completing the low pass filtering, the interpolation delay and the phase rotation which are matched with the signals at one time to realize the actual formation of delayed beams. As to the executive mode of wave filtering, a multi-channel wave filter with a cascade structure is disclosed. The wave filtering in the same path signal is completed in different filtering submodules by cascading while the different path signals share the same filtering module. The filtering output is completed with the sample selection, and the method of the invention is suitable for being realized in an ASIC or an FPGA and saves a great quantity of logic resources compared with the conventional wave filters.