57 results about "Outer product" patented technology

The invention discloses a bilinear feature fused fine-grained concept model and a learning method. The method comprises the following steps: carrying out external product, pooling and other methods onmulti-layer features in a fine-grained image extracted by a deep convolutional network model vgg16 to obtain bilinear feature descriptors with discriminative local features, fusing the descriptors, and fusing the multi-window features extracted by the feature graphs fused with VGG-16conv5_1 path, the conv5_2 path and the conv5_3 path with bilinear features extracted by the VGG-16conv5_2 path andthe conv5_3 path and sending the data to a full connection layer, and then connecting the data with a softmax multi-class classifier to obtain a classification result. In the data preprocessing stage,input image data are preprocessed, the image mean value is subtracted to eliminate noise, and data enhancement means such as image random clipping and image horizontal overturning are effectively utilized. Under the condition that only the class information of the image level needs to be provided, the classification precision is improved through fine-grained image multi-layer feature fusion.

A native tensor processor calculates tensor contractions using a sum of outer products. In one implementation, the native tensor processor preferably is implemented as a single integrated circuit and includes an input buffer and a contraction engine. The input buffer buffers tensor elements retrieved from off-chip and transmits the elements to the contraction engine as needed. The contraction engine calculates the tensor contraction by executing calculations from an equivalent matrix multiplications, as if the tensors were unfolded into matrices, but avoiding the overhead of expressly unfolding the tensors. The contraction engine includes a plurality of outer product units that calculate matrix mutiplications by a sum of outer products. By using outer products, the equivalent matrix multiplications can be partitioned into smaller matrix multiplications, each of which is localized with respect to which tensor elements are required.

Systems, apparatuses, and methods for implementing a low power parallel matrix multiply pipeline are disclosed. In one embodiment, a system includes at least first and second vector register files coupled to a matrix multiply pipeline. The matrix multiply pipeline comprises a plurality of dot product units. The dot product units are configured to calculate dot or outer products for first and second sets of operands retrieved from the first vector register file. The results of the dot or outer product operations are written back to the second vector register file. The second vector register file provides the results from the previous dot or outer product operations as inputs to subsequent dot or outer product operations. The dot product units receive the results from previous phases of the matrix multiply operation and accumulate these previous dot or outer product results with the current dot or outer product results.

The present invention discloses an image processing method, apparatus, storage medium and device, and belongs to the field of deep learning. The method comprises: for each convolution layer of a preset convolutional neural network, acquiring a feature map input in the convolution layer; performing first pre-processing on the feature map, and generating a first matrix according to the pre-processedfeature map, wherein feature data continuously used in the pre-processed feature map is arranged adjacent to the first matrix; performing second pre-processing on the weight of at least one convolution kernel of the convolution layer, and generating a second matrix according to the pre-processed weight, wherein the weights continuously used in the pre-processed weight are arranged adjacent to thesecond matrix; and performing an outer product operation on each row element in the first matrix and the second matrix, and after performing third pre-processing on the obtained outer product operation result, obtaining a convolution operation result output in the convolutional layer. According to the technical scheme of the present invention, in the implementation of the winograd convolution acceleration operation, the computational density and memory access efficiency are effectively improved, and the complexity of hardware implementation is reduced.

A method of improving accuracy and reliability of motion estimation is described herein. In one aspect, a 2D neighborhood of phase correlation peak is approximated with an outer-product of two 1D vectors to eliminate the sub-pixel error. In another aspect, estimation of reliability is improved. In yet another aspect, two-pass phase correlation is implemented to eliminate sub-pel motion bias.

In an embodiment, an outer product engine is configured to perform outer product operations. The outer product engine may perform numerous multiplication operations in parallel on input vectors, in an embodiment, generating a resulting outer product matrix. In an embodiment, the outer product engine may be configured to accumulate results in a result matrix, performing fused multiply add (FMA) operations to produce the outer product elements (multiply) and to accumulate the outer product elements with previous elements from the result matrix memory (add). A processor may fetch outer product instructions, and may transmit the instructions to the outer product engine when the instructions become non-speculative in an embodiment. The processor may be configured to retire the outer product instructions responsive to transmitting them to the outer product engine.

A native tensor processor calculates tensor contractions using a sum of outer products. In one implementation, the native tensor processor preferably is implemented as a single integrated circuit andincludes an input buffer and a contraction engine. The input buffer buffers tensor elements retrieved from off-chip and transmits the elements to the contraction engine as needed. The contraction engine calculates the tensor contraction by executing calculations from equivalent matrix multiplications, as if the tensors were unfolded into matrices, but avoiding the overhead of expressly unfolding the tensors. The contraction engine includes a plurality of outer product units that calculate matrix mutiplications by a sum of outer products. By using outer products, the equivalent matrix multiplications can be partitioned into smaller matrix multiplications, each of which is localized with respect to which tensor elements are required.

The invention discloses a tensor CP decomposition implementation method based on a distributed environment. A factor matrix A(n) in each iteration is updated based on an ALS algorithm, Y=X(n)(A(N)...A(n+1)A(n-1)...A(1)) is first calculated by splitting a Khatri-Rao product, then an outer product is calculated by parallel calculation, finally, a matrix Y and a matrix V are partitioned, partitionedmatrices corresponding to the matrix Y and the matrix V are distributed to a host of a Spark cluster by a Map operation, a matrix multiplication is performed by a Reduce operation, and then the multiplication results are sent to the host by the Map operation and merged by the Reduce operation to obtain A(n)=YV. The method implements tensor CP decomposition based on MapReduce and Spark technologies, thereby effectively improving the efficiency of tensor CP decomposition.

The invention discloses a signal number estimation method applied in incidence of incoherent signals and multiple groups of coherent signals on the basis of uniform linear arrays. After an outer-product matrix based on a cross covariance matrix of received signals of two linear arrays is acquired, a first combined matrix is constructed by the cross covariance matrix and a transformation matrix thereof, and an outer-product matrix of the combined matrix is acquired, the number of the incoherent signals and the number of the groups of coherent signals can be acquired according to the rank of the two outer-product matrixes. Further, a new oblique projector is estimated to suppress the incoherent signals of data of the received array, a series of cross covariance matrixes composed of data of one uniform linear array and a series of sub-arrays of the other uniform linear array form a new combined matrix, and the rank of the outer-product matrix of the combined matrix is equal to the number of the coherent signals. According to great quantities of experiments, the signal number estimation method with less snapshots and low signal-to-noise ratio is superior to the MDL/AIC method, the MENSE method and the SRP(smoothed rank profile test) which are subjected to the FBSS (front-rear space smooth) preprocess.

The invention discloses an adaptive estimation method for a spatially non-uniform sea clutter covariance matrix. The method comprises the following steps of first performing NSCM (normalized sample covariance matrix) estimation by utilizing a reference sample to obtain an NSCM estimation value, and preprocessing the NSCM estimation value; then determining a normalized outer product matrix of the reference sample, and preprocessing the normalized outer product matrix; next calculating a weighting coefficient by utilizing preprocessing data obtained by the two preprocessing; and finally substituting the weighting coefficient into an estimation algorithm for a diagonally loaded adaptive covariance matrix to obtain a final covariance matrix estimation value. According to the method, the aim of adaptively reflecting statistical properties of sea clutters is fulfilled, the influence of clutter power on the performance of a detector is eliminated, and the problem of finite usable reference samples is solved; and the method can be applied to covariance matrix estimation under a non-uniform sea clutter background.

To suitably reduce data amount. For example, feature points from a branch point or an end point to the next branch point or an end point in a blood vessel line are set as a group. In the three feature points satisfying one of the condition that the absolute value of the outer product of vectors in continuous three feature points is smaller than an outer product threshold value, and the condition that a cosine in the above three feature points is s maller than a cosine threshold value, the middle one of the

A dual receptacle aerosol sprayer (10) with a thin, flexible plastic outer receptacle (11) for product (12) and a substantially rigid inner receptacle (13) for propellant (14) seated within the outer receptacle (11). A closure (15) closes the inner receptacle (13) and contains a valve assembly (18). Primary propellant and secondary product valves (70, 80) in the valve assembly (18) control flow from the inner and outer receptacles (11, 13) up propellant and product valve stem bores (42, 43) into an actuator (21) having an aspirating nozzle insert (29) with a Venturi constriction (30). A conduit (22) extends from the valve assembly (18) through the inner receptacle (13) and into the outer receptacle (11). To avoid propellant overloading and rupture of the outer receptacle (11), a one-way tertiary valve (31) downstream of the secondary product valve (80) closes on clogging of the actuator discharge outlet (32) to prevent, during actuation, misdirected propellant flow from the actuator (21) and through the stem product bore (42) and the secondary product valve (80), from entering the outer product-containing receptacle (11).

The invention relates to an irregular area automatic matching method in a digital image, comprising the following steps: acquiring an image and inputting to a computer; extracting irregular areas in the image; determining the main direction of each irregular area through calculating the average gradient direction of points in the irregular area; determining the maximum symmetric position by calculating the asymmetry of each point; determining the neighbor area of each irregular area by taking the maximum symmetric position as the center and the main direction as the starting direction; dividing the neighbor area of each irregular area into three layers from inside to outside, and further dividing each layer into at least one subarea along peripheral direction; utilizing gradient inner product and outer product operations to construct descriptors matched with the irregular areas; and carrying out area matching and outputting results through calculating the Euclidean distance among the obtained descriptors and utilizing a bidirectional matching criterion to calculate the similarity among the irregular areas. The method of the invention cannot introduce error brought by fitting the shapes of the irregular areas into regular shapes, thus improving matched accuracy.

The invention relates to a bank bill image classification method based on improved B-CNN. The bank bill image classification method based on improved B-CNN comprises the steps: extracting position information of all information areas in a bill image; intercepting the bill image according to the position information to obtain a plurality of target image blocks; and sequentially inputting the plurality of target image blocks into an improved B-CNN model to perform feature extraction, feature cross fusion and feature outer product operation so as to realize classification of the bill image. According to the embodiment of the invention, the bank bill image classification method can achieve the classification of fine-grained images through employing the improved B-CNN model, can extract the convolution features with higher discrimination, can achieve the classification of different types of bill images with high similarity, and guarantees the higher classification accuracy.

The invention relates to the visual technical field of electromagnetic environments, and discloses an electromagnetic environment complexity quick mesh generation and display method. The method includes the steps that scatter points are distributed according to the ascending order or the descending order of x coordinates, and a first triangle is built, so that the triangle is large enough to contain all boundary points and the scatter points; Delaunary triangle positioning is carried out on sampled data through a vector outer product method; the triangle is updated through a vector included angle formula; equivalent boundary searching, equivalent boundary determining and isoline tracking are carried out on a triangle grid to obtain an isogram; electromagnetic nephogram rendering is conducted to obtain an electromagnetic environment grade display image, and smoothing processing is carried out through a cubic b-spline algorithm. The method can accurately master an electromagnetic complexity change tendency, is little in calculation quantity, can meet the requirement for triangle grid mesh generation and electromagnetic nephogram rendering of electromagnetic environment real-time dynamic drawing, and facilitates quick analysis of area electromagnetic environments.

The invention provides an object classification method based on an image. The method comprises the steps of obtaining a to-be-detected image with at least one to-be-classified object and the background, separating the to-be-classified object in the to-be-detected image from the background, extracting at least one feature of the to-be-classified object under multiple scales, performing channel superposition on the at least one feature, obtaining at least two channel superposition images, combining the at least two channel superposition images through the matrix outer product operation, obtaining a feature enhancement graph of the at least one to-be-classified object, classifying the feature enhanced graphs of the at least one to-be-classified object to obtain the classification result, so that the to-be-classified object is separated from the background, the classification efficiency of the to-be-classified objects in the subsequent steps is improved, the recognition and classificationaccuracy is improved, meanwhile, the recognition and classification accuracy and the classification efficiency of the densely-arranged objects with smaller distinguishable features are improved, and the training amount of a classification neural network is reduced.

A method and system for generating an output image from a plurality, N, of corresponding input image channels is described. A Jacobian matrix of the plurality of corresponding input image channels is determined. The principal characteristic vector of the outer product of the Jacobian matrix is calculated. The sign associated with the principal characteristic vector is set whereby an input image channel pixel projected by the principal characteristic vector results in a positive scalar value. The output image as a per-pixel projection of the input channels in the direction of the principal characteristic vector is generated.

The invention discloses a low-noise non-lens imaging method, which comprises the following steps: after a detector is powered on, recording light rays emitted by a target and passing through a coding template, and constructing a target image after an output pattern is subjected to coding pattern relevant processing, wherein effective area array size of the detector is larger than the pattern size of the coding template by half; the coding template is a coding pattern formed through etching based on etching technique after a metal film is plated on a glass substrate; the coding pattern is formed by a plurality light-transmitting and light-proof two kinds of same-size square units, the distribution of the square units being expressed by a matrix generated based on an outer product of two pseudo-random vectors formed by 0 and 1, the 0 and 1 in the matrix indicating light non-transmitting and light transmitting in the coding pattern respectively; and the number of the light-proof square units is larger than that of the light-transmitting square units. The method is based on a specific coding template, and can solve the problem that a non-lens camera based on a URA template is sensitive to noise, thereby improving signal-to-noise ratio of the reconstructed image.

The present invention provides a method and a system for detecting the intersection of cracks in a three-dimensional discrete fracture network of a tight oil-gas reservoir. The method comprises: for each crack polygon, generating a corresponding bounding box, filtering and obtaining intersected bounding box pairs through a projection method and an intersection algorithm, and initially determiningintersected crack pairs; eliminating a part of non-intersected crack pairs in the initially determined intersected crack pairs by using the inner product outer product method to obtain rough refined intersected crack pairs; and through the intersection point position relationship of two crack polygons, obtaining finely refined intersected crack pairs, and determining the intersection of the intersected crack pairs according to the intersection point position relationship. According to the method and the system provided by the present invention, by using the bounding box, the projection method,the intersection algorithm and the like, intersected crack pairs are initially determined, and obvious non-intersected crack pairs are filtered out, so that for understanding the intersection of thecracks in the three-dimensional discrete crack network of the tight oil-gas reservoir, a large quantity of computing time is saved, and the detection efficiency of the intersection of cracks is improved.

The invention relates to a die casting material with a plating layer, and a hole patching method of the die casting material. The method comprises the following steps: the die casting material with the metal plating layer is insulated and cooled; metal for forming the metal plating layer is tin and/or lead; and the insulation temperature is higher than the melting point of the metal plating layer, and is lower than the melting point of the die casting material. The invention further relates to the die casting material manufactured by the method, a die casting material surface electroplating method and a die casting outer part electroplated by the method. The die casting material obtained by the hole patching method prominently reduces the porosity, and forms a compact, complete, uniform and continuous layer on the surface; and the hole-patched die casting material is electroplated, so that the porosity of the outer part with the plating layer can be further reduced, the problem of a die casting material with defects in the prior art unsuitable for serving as an outer part is solved, and the application range of the die casting material to outer products is widened.

A native tensor processor calculates tensor contractions using a sum of outer products. In one implementation, the native tensor processor preferably is implemented as a single integrated circuit andincludes an input buffer and a contraction engine. The input buffer buffers tensor elements retrieved from off-chip and transmits the elements to the contraction engine as needed. The contraction engine calculates the tensor contraction by executing calculations from equivalent matrix multiplications, as if the tensors were unfolded into matrices, but avoiding the overhead of expressly unfolding the tensors. The contraction engine includes a plurality of outer product units that calculate matrix mutiplications by a sum of outer products. By using outer products, the equivalent matrix multiplications can be partitioned into smaller matrix multiplications, each of which is localized with respect to which tensor elements are required.

The invention provides a cross-domain heterogeneous large data multi-view clustering method and device based on tensor. The method includes: constructing sample tensor according to fused cross-domainheterogeneous feature space, and constructing feature space combination vector according to different context; accumulating the sample tensor to obtain a combined tensor; normalizing the order corresponding to each characteristic space of the merging tensor to obtain a connection tensor; obtaining the feature space scoring vectors according to the connection tensor, and obtaining a scoring tensorby outer product of the feature space scoring vectors; introducing the feature space combination vector and the scoring tensor into a high-dimensional space tensor distance to construct a combined scoring tensor distance; calculating sample similarity according to the combined scoring tensor distance, and constructing a view matrix according to the feature space combination vector; obtaining the multi-view clustering results under different views according to the view matrix. The invention solves the technical problem that in the prior art, different clustering results cannot be generated according to the requirements of different applications in different contexts, thereby providing high-quality clustering services for different applications.

The invention belongs to the relevant technical field of path generation, and discloses a thin-wall part robot machining path generation method based on machining parameters. The method comprises thefollowing steps of (1) uniformly dispersing the boundary curve of a thin-wall part to be processed into original path points; (2) respectively calculating to obtain a normal vector and a tangent vector of each original path point; (3) wherein the normal vector and the tangent vector are orthogonal; (4) calculating the outer product of the normal vector and the tangent vector to obtain the outer offset vector of the current path point, and establishing a moving coordinate system of the path point based on the normal vector, the tangent vector and the outer offset vector; (5) smoothing the direction of the normal vector based on the moving coordinate system, and simultaneously adjusting the moving coordinate system; (6) smoothing the outer bias vector based on the corner direction and the corner size of the outer bias vector; and (7) performing external offset on the path point to obtain a robot machining path of the thin-wall part to be processed. The thin-wall part robot machining pathgeneration method based on machining parameters is easy to implement and has stronger applicability.

The invention relates to a method and detection kit for detecting a helicobacter pylori drug resistance gene. The method comprises the following steps: carrying out PCR reaction on to-be-detected sample genome DNA serving as a template by using nested PCR outer primers; carrying out PCR by using the outer product of nested PCR as a template and using allele specific primers as inner primers; carrying out electrophoresis on the product (obtained after inner PCR amplification) in 1.5% agarose gel containing ethidium bromide; detecting whether a specific amplification strip exists or not under an ultraviolet ray transilluminator, wherein the nested PCR outer primer sequences are as shown in SEQ ID NO.1 and 2; the inner primer sequences are as shown in SEQ ID NO.3-9; and the kit comprises the nested PCR outer primers and inner primers. According to the method, the target of distinguishing whether drug resistance mutation sites exist in different specimens can be achieved, and the method is high in detection rate, high in sensitivity, simple and convenient to operate, cannot cause pain, is of a noninvasive type and has good patients' compliance.

The invention is directed to a system and methods of capturing rapid changes in complex systems by examining the interaction of the various components within the system. This may be accomplished by decomposing a multivariate signal into a dictionary of elements and a corresponding sparse mixing matrix. Each dictionary element may provide a rank-1 estimation of a disjoint partition of time points of the signal that may permit estimation of the covariance of each partition using an outer-product of the corresponding dictionary element.