31 results about "Similar distribution" patented technology

A computer-implemented method executed by at least one processor for training fast models for real-time object detection with knowledge transfer is presented. The method includes employing a Faster Region-based Convolutional Neural Network (R-CNN) as an objection detection framework for performing the real-time object detection, inputting a plurality of images into the Faster R-CNN, and training the Faster R-CNN by learning a student model from a teacher model by employing a weighted cross-entropy loss layer for classification accounting for an imbalance between background classes and object classes, employing a boundary loss layer to enable transfer of knowledge of bounding box regression from the teacher model to the student model, and employing a confidence-weighted binary activation loss layer to train intermediate layers of the student model to achieve similar distribution of neurons as achieved by the teacher model.

The present invention provides a system, method, and computer-readable medium for identifying and removing active malware from a computer. Aspects of the present invention are included in a cleaner tool that may be obtained automatically with an update service or may be downloaded manually from a Web site or similar distribution system. The cleaner tool includes a specialized scanning engine that searches a computer for active malware. Since the scanning engine only searches for active malware, the amount of data downloaded and resource requirements of the cleaner tool are less than traditional antivirus software. The scanning engine searches specific locations on a computer, such as data mapped in memory, configuration files, and file metadata for data characteristic of malware. If malware is detected, the cleaner tool removes the malware from the computer.

The invention relates to a vehicle type recognition method based on a CNN and domain adaptive learning. The method comprises steps of: establishing a CNN-based initial model by adding a rotation-invariant layer in an Alexnet network, distinguishing a discriminant layer and designing a new objective function; separately extracting the feature maps of different-domain sample convolution layers by using the established initial model, calculating the cosine similarity between the sample feature maps, determining the shared convolution kernel or the non-shared convolution kernel of the CNN, retaining the weight and the offset of the shared convolution kernel, and updating the weight and the offset of the non-shared convolution kernel; based on a target-domain training sample, calculating the cosine similarity between respective feature map layers and the average similarity of the entire target domain, and clustering each type of similar feature maps according to the average similarity; expanding source-domain samples with similar distribution characteristics in the target domain to new samples in the target domain, adjusting the entire CNN model by using the new samples in the target domain, and then using a softmax classifier to classify the vehicle types of the test samples in the target domain.

The invention discloses a remote sensing image ground object classification method based on superpixel coding and convolution neural network, using adaptive superpixel coding and double channel convolution neural network. The remote sensing image ground object classification method based on superpixel coding and convolution neural network includes the steps: utilizing a superpixel algorithm to perform image pre-segmentation; using a cluster method to merge neighboring and similar superpixel blocks, setting the size of the taken blocks, constructing three double channel convolution neural networks with different input size; inputting samples with different taken block size into the corresponding network; using the convolution neural networks to extract the data characteristics of two sensors respectively; merging the extracted characteristics for classification; and according to the size of the merged pixel block, determining the size of the taken blocks of the samples, and realizing adaptive selection of the utilized neighborhood information. The remote sensing image ground object classification method based on superpixel coding and convolution neural network can realize adaptive selection of the utilized neighborhood information to enable the neighborhood information to realize positive feedback effect and preferably utilize the neighborhood information to send the samples to different networks according to the neighborhood information so as to enable the samples with similar distribution to enter the same network, thus effectively improving the classification accuracy.

Method and arrangement are disclosed for a power distribution network for distributing power to VCSELS on an optoelectronic chip that maintains a constant bias voltage to the VCSELs. More particularly, the power distribution network includes an H-tree network for directing a power signal. The H-tree network includes a primary input for receiving the power signal and at least one level, with terminal nodes coupled to the last level, and each level includes a plurality of segments. Each segment of a respective level is equal in length and the total number of segments from the primary input to each terminal node is the same. Since the distance from the primary input to each terminal node is equal, greater bias voltage uniformity is provided to the terminal nodes. A similar distribution network is used for the ground connection.

The present invention relates to glial cell line-derived neurotrophic factor (GDNF), a potent neurotrophin that exhibits a broad spectrum of biological activities on a variety of cell types from both the central and peripheral nervous systems. The present invention involves the cloning and characterization of receptors for GDNF. Nucleic acid and amino acid sequences are described for GDNFR protein products. A hydrophobic domain with the features of a signal peptide is found at the amino terminus, while a second hydrophobic domain at the carboxy terminus is involved in the linkage of the receptor to the cell membrane. The lack of a transmembrane domain and cytoplasmic region indicates that GDNFR requires one or more accessory molecules in order to mediate transmembrane signaling. GDNFR mRNA is widely distributed in both nervous system and non-neural tissues, consistent with the similar distribution found for GDNF.

The invention discloses a ganglion differentiation-based incremental target identification system. The system is composed of a low hidden layer, a ganglion layer, a high hidden layer and a classifier, wherein the ganglion layer is located between the low hidden layer and the high hidden layer; ganglions of the ganglion layer extract a sample characteristic law, and through activation and differentiation of the ganglions, corresponding independent characteristic sets are formed in the high hidden layer and characteristic memory of samples is formed; a quantity of the characteristic sets of the high hidden layer is variable; newly added ganglions form new memory for new samples, and the characteristic sets of the high hidden layer are adaptively updated, so that incremental target identification is realized; the ganglions refer to neural network nodes used for representing a group of samples with similar distribution laws, and when an input sample characteristic parameter is greater than an activation threshold, the ganglions are activated and form the independent characteristic sets in the high hidden layer; when the activity of the activated ganglions is lower than a threshold, the ganglions die; and different ganglions are activated for the samples with different characteristics, and ganglions are newly added.

A failure analysis method according to the invention includes inputting the positions of failures in multiple wafers of an input device; preparing multiple sections in the multiple wafers; calculating feature amounts, which are represented by at least one numerical value representing a distribution of the failures in the multiple wafers, for each of the multiple sections; and representing by a first numerical value, the degree of similarity between the multiple wafers in terms of the feature amounts. Subsequently, the method includes detecting another wafer, which has the first numerical value greater than a predetermined first threshold, for each of the multiple wafers and forming a similar wafer group of multiple wafers with similar distributions of the failures.

The invention discloses a model updating device and method, a data processing device and method, for updating a target model in a multi-model system, wherein each model in the multi-model system is obtained by pre-training a training dataset in different ways. The model updating device comprises a pseudo label obtaining unit, a first feature distribution obtaining unit, a second feature distribution obtaining unit, an adjustment unit and an updating unit, wherein the pseudo label obtaining unit is used for using a calibration model to process a dataset to be tested and taking the processing result as a pseudo label; the first feature distribution obtaining unit is used for obtaining the feature distribution of the dataset to be tested on the basis of the pseudo label; the second feature distribution obtaining unit is used for obtaining the feature distribution of a training dataset based on a target model; the adjustment unit is used for adjusting the feature space subdivision of the target model according to the feature distribution of the training dataset and the feature distribution of the dataset to be tested, so that the training dataset and the dataset to be tested have similar distributions specific to the feature space subdivision; the updating unit is used for updating the target model by adopting the training dataset based on the adjusted feature space subdivision.

This invention relates to a method for packing large file service light transmitted self-similar cluster packets characterizing in taking if the time interval for arriving network edge nodes by data packets and/or packet length obey the self-similar distribution as the judgment basis to divide the accumulated data packets into CLUSTER and NON-CLUSTER packets and put them in different buffer storages to be sent out. This invention fully considers the accumulation method when services arrive at the network edge nodes to obey three kinds of self-similar math models and fully utilizes the character of the transmission of large files of flow media, HDTV and GRID network application to accumulate a large service flow accessing a pair of same entrance and exit nodes in a short period a CLUSTER and an independent data packet a CLUSTER and uses time threshold to meet the need of service to transmission delay to reduce the reconfiguration rate of network resources.

The invention discloses a metering device clock error trend prediction method based on a time sequence evolution gene model, and relates to the field of electric power operation and maintenance. Existing work can only solve the clock error problem of a specific scene and is not universal enough. According to the invention, a time sequence evolution gene model is adopted, the time sequence evolution gene model divides an ammeter clock error into a plurality of sub-sequences on a certain window, the model can analyze the characteristics of window sub-sequence error change, and the sub-sequenceswith similar distribution are divided into blocks through a classifier of the model; through the generative adversarial network, genes generating the subsequence distribution characteristics are mined; and combining genes of each subsequence in history, analyzing an evolution process through a recurrent neural network, analyzing an evolution mode of the subsequence, and predicting a future clock error trend of the subsequence. The technical scheme is not limited to specific devices and causes of errors, and can be adaptive to different scenes, so that a more universal scheme is provided, and the clock error problem of the generalized electric energy device is solved.

The invention relates to a power system operation state simulation method and a power system operation state simulation system. The method comprises the following steps of: performing statistical analysis on the basis of historical data of the operation state of a power system; obtaining probability distribution curves of three kinds of random variables including the power generator active power output, the power generator voltage and the active load of a load node of a power generator; and using a clustering analysis method for respectively clustering the probability distribution curves with similar distribution in the three kinds of random variables. The dimension reduction goal is achieved; the time period for generating mass power network operation state simulation samples can be greatly shortened; and the efficiency is improved. The clustering probability distribution curves subjected to clustering merging dimension reduction are sampled; the power flow distribution is subjected to simulation calculation; a power network operation state simulation sample is obtained; the obtained power network operation state simulation sample is similar to the power network actual operation state; the problem of low precision of the calculation result due to random sampling of each random variable in a designated range in the existing method is effectively solved; and the precision of the power flow distribution calculation result can be effectively improved.

An MS² analysis for one precursor ion is performed to collect data on each micro area within a measurement target area (S1). A plurality of product ions are extracted based on those data (S2), and a mass spectrometric (MS) imaging graphic is created for each m/z of the product ion (S3). Hierarchical cluster analysis is performed on the created MS imaging graphics to group the product ions based on the similarity of the graphics (S4). Product ions having similar distributions are sorted into the same group. Such a group of ions can be considered to have originated from the same compound. Accordingly, the intensity information of a plurality of product ions is totaled in each group and for each micro area (S5), and an MS imaging graphic is created based on the totaled intensity information (S6). Even if there are a plurality of compounds overlapping the precursor ion, the influence of the overlapping can be eliminated through those steps. Thus, a graphic having a higher level of SN ratio, sensitivity and dynamic range than an MS imaging graphic obtained at a single product ion can be created and displayed.

The invention discloses a system reliability test method and a system reliability test device. The system reliability test method mainly includes: configuring multiple reliability sets for each node in a system to be tested through a random mode based on a minimal path set method, confirming reliability degrees of the system through a simulation test, comparing a distribution state of a plurality of the confirmed reliability degrees of the system with distribution states of various preset distribution types after all the reliability sets are completely executed, and using the preset distribution type of the distribution state most similar to the distribution state of the reliability degrees of the system as the reliability distribution type of the system. Accordingly, the system reliability test method and the system reliability test device avoid the problems that calculated amount of the reliability degrees of the system is huge due to the fact that complexity of the system is increased, and then complexity of a reliability test is increased in the prior art, and simultaneously avoid uncertainty of a comprehensive test of the system.

The invention discloses an SILK security steganography method based on LSF coefficient statistical distribution characteristics. The method comprises the steps: firstly, carrying out the statistical analysis of the code word usage condition of an LPC quantization stage in the SILK coding process, and carrying out the grouping of groups with similar code word distribution conditions according to the statistical condition; dividing code words with similar distribution in each level in the LSF quantization process into one group; finally, adopting the LSF indexes of the first level to the ninth level of an obtained quantization result for embedding the secret information, and determining the bit number of the embedded secret information by the number of code word vectors in an embedded group.According to the SILK-oriented safe steganography method, the SILK-oriented safe steganography algorithm based on LSF coefficient statistical distribution characteristics is achieved for the problemthat an existing SILK-oriented steganography method is relatively lacked. The maximum embedding capacity of the algorithm can reach 274 bps. Meanwhile, the auditory concealment of voices can be guaranteed, and the good safety is achieved.

The invention relates to an image processing apparatus and method. The image processing apparatus includes a converter that converts an input image into a depth image to obtain a converted depth for each pixel of the input image; and a classifier that classifies between the converted depth and a real depth from a depth data set, where the classifier and the converter are trained until the classifier cannot distinguish the converted depth from the real depth. By using the image processing apparatus and method according to the present disclosure, depth can be learned from a single image by confrontation training, which not only can learn the depth of the single image by a converter, but also can understand the high order consistency of the depth by a classifier. Through such adversarial training, the converter may output a depth map having a similar distribution as a real depth map.

The invention discloses a mathematical modeling method and system, equipment and a computer readable medium. The mathematical modeling method comprises the following steps: a strict variable screening step; a step of mining client setting information as variables; a step of screening the variables through set features of the variables; and a model training step. The external resource information characteristics are independently modeled, model results are fused in cooperation with models established by other characteristics, and when the external resource information has a problem, only the resource information model is adjusted without affecting the whole model; and products with similar distribution and scenes are combined together for modeling. According to the mathematical modeling method, the system, the equipment and the computer readable medium provided by the invention, the stability of the mathematical model can be improved, and meanwhile, the recognition efficiency and accuracy of the customer quality can be improved. According to the method, the instability of variables and the over-fitting of the model can be reduced, and the influence of data instability on model training and the final effect is reduced as much as possible.

The invention discloses a tread pattern structure of a fat tire. A plurality of raised pattern blocks are distributed at intervals on the tire tread. The pattern blocks are arranged symmetrically along the equator plane. The tread is divided into a main contact area in the middle and tire shoulder areas on both sides. In the main grounding area, there are horizontal stripes that are arranged on the equatorial plane and are similar to an inverted C shape. Vertical stripes are arranged between two adjacent horizontal stripes and located on both sides of the vertical stripes. The whole is an "S" shape with two oblique ribs, and irregular polygonal granular first shoulder patterns and second shoulder patterns are alternately distributed along the circumferential direction in the shoulder area. The invention adopts the mixed design of strip pattern blocks and granular pattern blocks to meet the comprehensive performance of fat tires in various complex road conditions such as snow fields, beaches, light forest roads, and off-road conditions.

The invention provides a reactive compensation differentiation collocation method of 10kV distribution lines. Firstly the 10kV distribution lines of a power grid are classified, and ground state models are constructed according to characteristic parameters of each kind of lines; then sensitivity between the characteristic parameters of the lines and a reactive compensation optimized collocation rate under the ground state models of each kind of the lines is calculated, reactive compensation collocation recommendation range tables under two-dimensional or multi-dimensional changes of the characteristic parameters of the lines are calculated, and a reactive compensation collocation rate recommendation mid-value general table of the lines is formed; and eventually according to the reactive compensation collocation rate recommendation mid-value general table of the lines, the optimized reactive compensation collocation for the 10kV distribution lines is guided. The reactive compensation differentiation collocation method of the 10kV distribution lines avoids problems of solution complication of a sensitivity coefficient and large calculated amount, and can visually and effectively confirm the reactive compensation collocation of similar distribution lines.

The invention discloses a random sampling method based on an information distribution mode, relating to a method for adaptively establishing a random sampling point that has similar distribution characteristics with an original signal distribution mode to perform original signal acquisition. The device adopted by the method mainly consists of a signal acquisition module (1), a signal sensor (5), an FIFO module (8), a random data generation module (9), a computer (10), a Y distribution module (12) and a signal analysis module (13). When the sound bandwidth is 22.05KHz, the Nyquist sampling rateneeds to take 44.1KHz, and the original sound can be completely reconstructed; however, when reconstructing the original sound by using normal distribution mode sampling, the sampling frequency onlyneeds to take 400Hz, and the sound effect with an error of less than 0.3% can be obtained; and in the case of obtaining the same signal resolution by using Y distribution mode sampling and Nyquist mode sampling, the data volume of Y distribution mode sampling can be greatly compressed.