48results about How to "Solve problems with high computational complexity" patented technology

The invention discloses a neural network black box aggressive defense method based on knowledge distillation, and the method comprises the steps: selecting a plurality of sub-networks to construct a teacher network, softening the input vectors of softmax layers of all sub-networks, and then reloading the model parameters of the sub-networks for training to obtain a new sub-network; obtaining a prediction label of each sub-network, and taking all prediction labels as soft labels after averaging or weighted averaging; inputting the Image Net data set into a student network, and guiding student network training by adopting a soft label, a data set hard label and a black box model special label to obtain a substitution model; adopting a white box attack algorithm to attack the substitution model to generate an adversarial sample sequence, adopting the adversarial sample sequence to attack the black box model, and selecting an adversarial sample successfully attacked in the adversarial sample sequence; and adding the successfully attacked countermeasure sample into a training set of the black box model, and performing countermeasure training by adopting the updated training set to generate the black box model with attack defense.

The invention relates to a log likelihood estimation based novel passive locating and tracking method. A multistation angle-measurement passive-location non-linear measurement equation is subjected to pseudo-linearization processing by the aid of log likelihood estimation, target position is obtained by means of maximum likelihood estimation, a closed-form solution of a target locating algorithm is given, then a 'current' statistical maneuver model and Kalman filtering are utilized to calculate target position, velocity and acceleration online in real time, and real-time accurate positioning and tracking on the target is realized.

The invention discloses an infrared moving small target detection method of a complex scene, which mainly comprises the following steps: 1) extracting a moving target area. 2) The image is divided into two parts: foreground and background according to the gray characteristics of the image. 3) Optical flow estimation is performed on the extracted moving target region by optical flow method, and thetarget motion vector information is extracted. 4) The extracted moving target region is morphologically filtered, the connected regions of the filtered moving target region are analyzed, and different moving target regions are identified. 5) the detection probability of the connected region is set, and the number of sample characteristic quantities to be detected in the connected region is determined. All connected domains are randomly sampled. The invention reduces the complexity of the algorithm, improves the detection accuracy rate, and effectively solves the problems of target misdetection, missed detection and mismatch caused by the large difference between the background and the target and the partial occlusion of the target.

The invention discloses a method and device for timing shift estimation. The method comprises a step of filling zero at two ends of a sequence p, so as to acquire a sequence s(k), wherein the sequence p is a result of conjugate multiplication between a frequency domain mark received on a preset channel by user equipment (UE) and a Zadoff-Chu sequence; k is equal to 0,...,M-1; M is the number of sub-carriers occupied by signals in the preset channel; and the preset channel comprises a PRACH (physical random access channel) or a PUSCH (physical uplink shared channel). The method further comprises the following steps: converting s(k) into a power sequence b(k); calculating the center of the power sequence b(k); and calculating the timing shift according to the expected peak position of UE and the center. By using the method and device provided by the invention, the complex rate of calculation is reduced.

The invention relates to a lane line detection method and system. The method comprises: acquiring an original image in front of a driving vehicle and determining a to-be-detected area from the original image; carrying out inverse perspective transformation on the image of the to-be-detected area; carrying out lane line enhancement processing on the image after the inverse perspective transformation; screening out candidate points from the image after enhancement; the screened candidate points are grouped; fitting is carried out on all candidate point groups to obtain fitting results being allto-be-detected lane lines; and carrying out inverse transformation on all to-be-detected lane lines obtained by fitting to original image space. According to the invention, after inverse transformation on the a to-be-detected area, lane line enhancement processing is carried out by using an image filter and rotary transformation is carried out on the image by combining principal direction estimation, so that the fluencies caused by lane changing and the like are eliminated and the robustness of the subsequent lane detection is improved. The image after enhancement processing is divided into aplurality of histograms, so that the computing load is reduced substantially by block division processing and the real-time performance of the system detection is enhanced.

Embodiments of the invention provide a bipartite graph-based service recommendation method and apparatus. The method comprises the steps of acquiring a service which each target user is interested in; constructing an initial user interest vector of each target user and an initial service topic vector of each service, wherein the user interest vector and the service topic vector are both probability vectors; generating a bipartite graph by taking the target users and the services as vertexes, and constructing a bipartite graph-based target function P according to the user interest vectors and the service topic vectors; solving the target function P to determine an element value of each user interest vector and an element value of each service topic vector; and performing service recommendation on the target users according to the user interest vectors and the service topic vectors. By applying the method and the apparatus, the calculation complexity of bipartite graph-based service recommendation is lowered, and the user interest vectors and the service topic vectors output by a model are all the probability vectors, so that more probability-based actual applications can be constructed conveniently.

The invention discloses an attribute alignment method and apparatus, relates to the field of data analysis, and aims to solve the problem of high calculation complexity caused by matching between an attribute of a data source and multiple attributes of another data source. The attribute alignment method comprises the steps of obtaining N data sources; according to the N data sources, obtaining anattribute name set and generating an internal dictionary; according to the attribute name set, the internal dictionary and an attribute name of each data source in the N data sources, obtaining an attribute eigenvector of each data source; according to the attribute eigenvector of each data source, calculating the similarity between any two data sources in the N data sources and a similarity matrix; according to the similarity matrix, clustering the N data sources to obtain k clusters; and according to the internal dictionary, performing attribute alignment on the data sources of the same cluster in the k clusters to obtain k data sources, and performing attribute alignment on the k data sources to obtain a result data resource. The attribute alignment method and apparatus is applied to the data analysis.

The invention provides an incremental track anomaly detection method based on incremental kernel principle component analysis, and belongs to the field of an incremental track anomaly detection method. The method comprises the following steps: to begin with, carrying out model initialization calculation, carrying out initial kernel feature space calculation through conventional Batch KPCA, and when M newly-increased track data comes, carrying out standardization on the M track data first; then, calculating kernel feature space of the newly-increased data through Batch KPCA; calculating average reconstruction error of the newly-increased data and training data, and if the error of the two is larger than a preset threshold value, using a follow-up kernel feature space division-merging method to update kernel feature space; then, carrying out projection on the updated kernel feature space and extracting a principal component; and finally, carrying out unsupervised learning and anomaly detection by utilizing a support vector machine. The advantages are that the method is superior to a conventional kernel principle component analysis method; computing complexity is reduced; and track anomaly detection efficiency is improved.

The invention provides a Granger causality discrimination method based on a quantitative minimum error entropy criterion. According to the method, the coefficient and the order of a regression model are determined by adopting the quantitative minimum error entropy criterion and a Bayesian information criterion, a causality discrimination index is obtained by calculating the error entropy and coefficient, and the causality between two time sequences is determined according to a causality judgment standard. Compared with a traditional Granger causality discrimination method based on a minimum mean square error criterion, the method is more accurate in estimating coefficients of the regression model, the obtained error entropy is smaller, and the causality discrimination index can be more accurately calculated. Due to the adoption of a quantization method, the calculation complexity of the method is remarkably reduced. The method integrates the error entropy and the coefficient when calculating the causality discrimination index, which makes the calculation of the causality discrimination index more accurate and robust. Therefore, the Granger causality discrimination method based on the quantitative minimum error entropy criterion provided by the invention is more easily promoted and used in practical applications.

The disclosed embodiment discloses a method, a device, a computer device and a storage medium for determining a cluster-like center. The method comprises the following steps: adopting a geographic position coding technique, converting each two-dimensional geographic position information in a geographic position information set to be processed into one-dimensional geographic position coding information; A dictionary tree is generated according to the location coding information. A tree node in the dictionary tree corresponds to a set geographic location area, and a geographic location area corresponding to a child node belongs to a geographic location area range corresponding to a parent node of the child node. Calculating a density value corresponding to each geographic position information according to a positional relationship between each tree node and each geographic position information in the dictionary tree and a quantity value of geographic position information associated witheach tree node; At least one cluster center is determine in that geographic location information set based on the density value. The technical proposal of the disclosed embodiment can reduce the computational complexity of the cluster center in the clustering algorithm.

The invention provides a target tracking method and device, belonging to the field of computer vision and mode identification. The method and the device are used for solving the problem of high complexity of an existing target tracking algorithm. The method comprises the steps of selecting a first target position in an Ith frame, determining the tracking result of a first target as a first positive sample, determining the characteristic set of a first sample set from multiple first positive sample packets and first negative sample packets through a formula (1), determining a first MIL1-norm SVM through the characteristic matrix formula (2) of first training data, constructing a first minimum support sample set through a formula (3), determining the weight of the first minimum support sample set through a formula (4), determining the classification score of the first support sample set corresponding to the MIL1-norm SVM based on weight distribution through a formula (5), and classifying candidates through the first MIL1-norm SVM in an (I+1)th frame and determining the second target position in a (I+1)th frame image according to the weight of a second minimum support sample set and a formula (6).

The invention relates to the field of image segmentation, in particular to a self-adaptive image segmentation method based on an Otsu method and a K-means method. According to the method, the slack variable is introduced, the threshold value in the variable range is used as the local threshold value, the new threshold value search method is cyclically called by using the queue, and the plurality of local optimal threshold values are quickly and adaptively obtained, so that the problem of high calculation complexity of a multi-threshold Otsu method in an existing adaptive K-means image segmentation method is solved; the plurality of obtained thresholds are taken as the initial centroid of the K-means method, and the number of iterations of the K-means method is reduced; and the threshold obtained by clustering through the K-means method is used as the global optimal threshold, so that the image can be segmented accurately. According to the method, image illumination preprocessing servesas the purpose, the image can be segmented rapidly, accurately and adaptively, then the segmented area is corrected, and therefore information, lost due to illumination influence, of the image is recovered.

The invention discloses a multi-input multi-output visible light communication system based on coefficient separation. At a transmitting end, LEDs are divided into two groups. The first group of LEDs transmits the positive integer part of a bipolar pulse amplitude modulation signal, and the second group of LEDs transmits the absolute value of the negative integer part of the bipolar pulse amplitude modulation signal. At a receiving end, a novel receiver with low computing complexity is utilized; the receivers utilizes the characteristic of a transmitted signal; the absolute value of the bipolar pulse amplitude modulation signal is estimated; and then the polarity of the bipolar pulse amplitude modulation signal is estimated; and finally estimation of the bipolar pulse amplitude modulation signal is finished. The multi-input multi-output visible light communication system with separated positive coefficient and negative coefficient according to the invention has advantages of low computing complexity, high performance, etc.

The invention discloses an uncertain data classification method based on direct distinguishing sequence mining. For the uncertain data set UTD, firstly, a class label is initialized and a class labelset is given. Under the class label set, the minimum support threshold is derived based on the given information gain threshold. The pattern growth strategy is used to enumerate the sub-sequences, generate the pattern candidate x, mine the discrimination sequence, and adopt the reduction strategy to generate the final discrimination sequence result set Rs. Then, the result set Rs is checked, and the closed sequence detection algorithm is used to determine whether each candidate sequence in Rs is a probabilistic frequent closed sequence or not. If the sequence is probabilistic frequent closed sequence, the discriminant sequence satisfying the condition is added to the result set RsTmp. Finally, by combining with the rule-based classification method or support vector machine existing maturedata classification methods, the data classification is completed. As that complement of the discriminate pattern mining on the uncertain data set, the method of the invention remarkably improves theefficiency, and the result set is more concise.

The related minimal normalization image segmentation method comprises: 1) reducing dimension to the target image; 2) taking minimal normalization segmentation to obtain the low-dimension segmentation boundary; 3) taking additional sampling, reconstructing an image with same dimension as the target; 4) connecting the discontinuous segmentation boundary to obtain the required segmentation boundary. This invention has acceptable time and space complexity and wide application.