111 results about "Convolution theorem" patented technology

The invention discloses an image target recognition method based on optimized convolution architecture. The image target recognition method includes collecting and enhancing an input image to form a sample; training the sample on the basis of the optimized convolution architecture; performing classified recognition on an image target by using the convolution architecture after training, wherein optimization of convolution architecture includes ReLU activation function; locally responding to normalization; overlapping and merging a convolution area; adopting neuron connection Drop-out technology; performing heuristic learning. Compared with the prior art, the image target recognition method has the advantages that tape label samples can be expanded, and the image target recognition method is supportive of classification of many objects and acquiring of high training convergence speed and high image target recognition rate and has higher robustness.

The invention discloses a natural scene text identification method based on a convolution attention network; the method comprises the following steps: using a two-dimensional convolution CNN as an encoder, extracting High-level semantic features of an input image, and outputting a corresponding feature graph to a decoder; using a one-dimensional convolution CNN as the decoder, combining an attention mechanism to integrate the High-level semantic feature formed by the encoder with a character level language model, thus forming a decode character sequence corresponding to the input image. For asequence with the length of n, the method uses the CNN modeling character sequence with a convolution kernel as s, and only O (n/s) times of operation is needed to obtain the long term dependence expression, thus greatly reducing the algorithm complexity; in addition, because of convolution operation characteristics, the CNN can be better paralleled when compared with a RNN, thus performing advantages of GPU resources; more importantly, a deep model can be obtained via a convolution layer stacking mode, thus improving abstractness expression of a higher level, and improving the model accuracy.

The invention provides a neural network training method, a neural network training device, a data processing method and a data processing device. The neural network training method comprises the following steps: S210, transforming a set of initial convolution kernels corresponding to each of at least one set of convolution layers of a convolutional neural network into a corresponding set of transformed convolution kernels by use of a low-rank approximation method; S220, training the convolutional neural network based on the transformed convolution kernels corresponding to the at least one set of convolution layers; S230, judging whether the trained convolutional neural network meets a predetermined standard, going to S240 if the trained convolutional neural network meets the predetermined standard, or going to S250; S240, decomposing the product of the set of trained convolution kernels corresponding to each of the at least one set of convolution layers into a corresponding set of compressed convolution kernels; and S250, taking the set of trained convolution kernels corresponding to each of the at least one set of convolution layers as a set of initial convolution kernels corresponding to the set of convolution layers, and returning to S210. Through the methods and the devices, the amount of computation can be saved.

The invention discloses a neural network structured pruning compression optimization method for convolutional layers, and the method comprises the steps: (1), carrying out the sparse value distribution of each convolutional layer: (1.1) training an original model, obtaining the weight parameter of each convolutional layer capable of being pruned, and carrying out the calculation, and obtaining theimportance score of each convolutional layer; (1.2) according to the sequence of importance scores from small to large, carrying out average scale segmentation by referring to the maximum value and the minimum value, carrying out sparse value configuration from small to large on the convolution layers of all the sections in sequence, and through model training adjustment, obtaining sparse valueconfiguration of all the convolution layers capable of being pruned; (2) structured pruning: selecting a convolution filter according to the sparse value determined in the step (1.2), and carrying outstructured pruning training; Wherein only one convolution filter is used for each convolution layer. According to the optimization method provided by the invention, the deep neural network can be more conveniently operated on a resource-limited platform, so that the parameter storage space can be saved, and the model operation can be accelerated.

The invention belongs to the technical field of neural networks, and relates to an FD-SOI process-based calculation accelerator in binary convolutional neural network memory. The accelerator utilizesthe back gate voltage of the FD-SOI-MOSFET to adjust its threshold voltage to achieve XOR processing of the data. The convolution kernel parameters of the convolutional neural network are "one-dimensional" processed and stored in the memory, and the convolution process of the convolution kernel to the neural network is realized by XOR operation of the convolution kernel by the FD-SOI-MOSFET. Underthe premise of in-memory calculation, compared with the traditional convolution process, the XOR operation is completed by using the XOR operation while maintaining high precision, which greatly improves the convolution processing speed of the neural network and saves the neural network. Parameter storage space, data transfer, and reduced computing power.

The invention provides an image classification method and device, and relates to the field of artificial intelligence, in particular to the field of computing vision. The image classification method comprises the following steps: obtaining a convolution kernel parameter of a reference convolution kernel of a neural network and a mask tensor of the neural network, and performing Hadamard operationon the reference convolution kernel of the neural network and the mask tensor corresponding to the reference convolution kernel to obtain a plurality of sub-convolution kernels; and performing convolution processing on the to-be-processed image according to the plurality of sub-convolution kernels, and classifying the to-be-processed image according to the convolution feature map finally obtainedby convolution to obtain a classification result of the to-be-processed image. Due to the fact that the mask tensor occupies a smaller storage space relative to the convolution kernel, some devices with limited storage resources can deploy a neural network comprising the reference convolution kernel and the mask tensor, and then image classification is achieved.

The embodiment of the invention discloses a method, device and system for data processing. The method comprises the steps of obtaining input data; obtaining convolution operation characteristics of afirst convolution layer, wherein the first convolution layer is any convolution layer in a neural network model, the neural network model comprises at least one convolution layer, and the convolutionoperation characteristics comprise at least one of the following parameters: the size of a convolution kernel, the moving step of the convolution kernel, the number of input channels and the parallelgranularity of a processor; obtaining a plurality of data blocks with the number being the same as the parallel granularity of the processor from the input data according to the convolution operationcharacteristics, wherein each of the plurality of data blocks is used for performing convolution processing with the convolution kernel; and performing convolution processing on the plurality of datablocks in parallel so as to obtain output data. The method, device and system disclosed by the embodiment of the invention can adaptively select an optimal parallel operation according to the convolution operation characteristics of different convolution layers, so that the parallelism of the convolution operation is improved.

The invention comprises the following steps: with the accumulating metrics of the initial window, an incorrect initial path is merged into a correct path, and a state of optimal node at the end of initial window is used as a decoding initial state of the tail-biting convolution code; in term of said decoding initial state and a delay window, through whole decoding trace-back, a Viterbi decoding is made for the tail-biting convolution code.

The invention relates to a convolution theorem based face verification accelerating method, and belongs to the field of face verification in computer vision. For a face verification system adopting a CNN (Convolutional Neural Network) technology, a convolution theorem method is adopted to replace the conventional convolution computation method to perform convolution computation on a convolution layer meeting an acceleration condition on the basis of using a GPU parallel computation platform. The convolution theorem shows that convolution in the space domain is equivalent to product in the frequency domain. Through transforming time-consuming convolution computation into product computation in the frequency domain, the computation amount can be significantly reduced, and the computation speed of a CNN is accelerated. In allusion to problems of great computational burden and slow operation speed of the face verification system, the method enables the operation speed of the face verification system to be obviously improved, and the processing capacity for mass data can be improved.

The present invention discloses a convolutional neural network construction method, and belongs to the neural network technology field. According to the present invention, when the convolutional neural network transfers forward, on each original convolution kernel, and by manually adjusting the dot product of a kernel and the original convolution kernel, the original convolution kernel is modulated to obtain the modulated convolution kernel, and the modulated convolution kernel is used to substitute for the original convolution kernel to transfer the neural network forward, thereby realizing afeature enhancement effect. According to the method of the present invention, the neural network is optimized greatly, so that the total number of the kernels that the network must learn is reduced.In addition, by modulating to generate the sub-convolution kernels to arrange the kernels of the redundant learning in an original network structure, a model compression purpose also can be achieved.

The present invention designs a convolutional neural network-based digital image steganalysis method. The convolutional neural network-based digital image steganalysis method comprises the following steps of S1 constructing a convolutional neural network formed by connecting multiple convolutional layers in series; S2 for the first convolutional layer, adopting a high-pass filter to initialize a convolution kernel, and then adopting a truncation linear unit activation function as the activation function of the convolutional layer; S3 inputting a digital image in the convolutional neural network, and using the convolutional neural network to output a result of the digital image whether or not after the steganography.

An iterative decode system and method for high-speed decode of Turbo code features that in iterative decoding procedure, the hard decision values for information bits and check bits are output, if they can meet the constrain relation of convolution code is checked, and the iteration is stopped if they can.

The invention discloses a partial binary convolution method suitable for embedded equipment, belonging to a model compression method of depth learning. The method comprises the following steps: 1. foreach convolution layer of a given CNN, measuring the importance of each convolution core according to the statistics of each output characteristic map; the convolution kernels of each layer being divided into two groups. 2, rearrangeing that two groups of convolution core on the storage space so that the storage positions of the same group of convolution cores are adjacent to each other, recording the rearrange order to generate a new convolution layer; 3, changing the channel order of the convolution nucleus of the next convolution layer according to the rearrangement order of the step 2; Step 4, according to the CNN processed in the above steps, fine-tuning training being carried out, binary quantization being carried out on the convolution layer divided into non-important layers, and the accuracy of the whole network being gradually recovered through the iterative operation of quantization and training. The invention ensures the accuracy of the given CNN on the large data set, andreduces the calculation and storage overhead of the convolution neural network on the embedded equipment.

The present invention relates to a method for determining a matrix of transmission cross coefficients w for an optical modeling in an optical proximity correction of mask layouts. In a first step, there is calculation of Fourier transforms of an illumination aperture, a lens aperture and a complex conjugate lens aperture, which are present in the form of image matrices with a predetermined raster. A second step involves calculating Fourier transforms for the transmission cross coefficients w from the Fourier transforms by means of a convolution theorem in order to obtain the matrix of the Fourier transforms of the transmission cross coefficients w. A further step involves inverse-transforming the Fourier transforms of the transmission cross coefficients w by means of a fast Fourier transformation in order to obtain the matrix of the transmission cross coefficients w for the optical modeling in the optical proximity correction of mass layouts.

The invention discloses a universal convolution-pooling synchronous processing convolution kernel system, and belongs to the technical field of convolution neural network acceleration in machine learning. An existing machine learning method is realized by adopting software, so that the problems of limited computing power, higher cost and the like exist. According to the present invnetino, the machine learning is realized by adopting the hardware design, the purpose of accelerating the convolutional neural network is realized in a convolutional-pooling synchronous processing manner, and the machine learning can be realized quickly and efficiently with low power consumption on the premise that the accuracy is not changed. According to an existing convolutional neural network, a convolution kernel is generally of a fixed size and cannot adapt to various design requirements, but the convolution kernel of the invention can change the parameters such as the size and the step length of the convolution kernel and can adapt to the design requirements under various conditions.

The invention discloses a dilated-convolution method based on a deep convolutional adversarial network model, and belongs to the fields of deep learning and neural networks. The dilated-convolution method includes the following steps: S1, constructing an original generative adversarial network (GAN) model; S2, constructing deep convolutional neural networks to use the same as a generator and a discriminator; S3, initializing random noises, and inputting the same into the generator; S4, utilizing dilated convolution to carry out convolution operations on images in the neural network; and S5, inputting a loss function value, which is obtained by the dilated-convolution operations, into the generator for subsequent training. According to the dilated-convolution method based on the deep convolutional adversarial network model constructed in the invention, a convolution manner of the discriminator and the generator after receiving a picture is changed, the discriminator and the generator are enabled to learn features of the image in a larger range, and thus robustness of the entire network training model can be improved.

The invention relates to a convolutional code decode method, which comprises: based on Viterbi algorithm, obtaining globe optimized path, obtaining decode data; checking the decode data; if the code is right, outing code, ending; or else, searching for other globe optimized paths and decode data, as based on known globe optimized paths, separating from selected path at each time of each globe optimized path, searching for local sub-optimized path, obtaining the difference between relative globe optimized path and optimized path, obtaining one globe optimized path, obtaining relative decode data; checking the decode data, if the check is wrong, searching for other globe optimized paths.

The invention discloses a Winograd convolution splitting method for a convolutional neural network accelerator. The method comprises the following steps: 1) reading an input and a convolution kernel of any size from a cache of the convolutional neural network accelerator; 2) judging whether convolution splitting is carried out or not according to the convolution kernel size and the input size, andif convolution splitting needs to be carried out, carrying out the next step; 3) splitting the convolution kernel according to the size and the step length of the convolution kernel, and splitting the input according to the size and the step length of the input; 4) combining and zero-filling the split elements according to the size of the convolution kernel, and combining and zero-filling the split elements according to the input size; 5) performing Winograd convolution on each pair of split input and convolution kernels; 6) accumulating the Winograd convolution results of each input and convolution kernel, and 7) storing the accumulation results in a cache of the convolutional neural network accelerator, so that the convolutional neural network accelerator can support convolution of various different shapes by adopting one Winograd acceleration unit.

The invention discloses an unmanned aerial vehicle visual target tracking method based on scale self-adaptive kernel correlation filtering, which comprises the following steps of selecting a trackingtarget, calculating to obtain the color and gradient initial probability density of a first frame of the tracking target, and training a classifier and detecting the central position of the target byusing the kernel correlation filtering algorithm for the first frame of data; establishing a one-dimensional kernel correlation filter from the second frame to detect the change of the target scale, and calculating kernel correlation filtering by using a convolution theorem; constructing a similarity function by utilizing the current target feature and the initial feature, if the similarity is smaller than a set threshold value, considering that the target identification is inaccurate or the target is lost, entering global search, otherwise, representing that the target is identified and tracked, and obtaining target position information; and sending the position information of the tracking target to an unmanned aerial vehicle flight control system in real time to control the position of the unmanned aerial vehicle. According to the method, the problem of fixed tracking scale of a kernel correlation filtering algorithm is optimized, and the tracking precision of target characteristicsis effectively improved.

The present invention relates to a spatial layering disturbance gravitational field grid model rapid construction method which can effectively solve the problem of obtaining the disturbance gravitation of earth's any height level rapidly and at a high precision. The method comprises the following steps of modifying a conventional integration formula into a convolution form to calculate rapidly by utilizing a convolution theorem; calculating the low-order model gravity anomaly and the residual gravity anomaly to obtain the residual gravity anomaly; carrying out the Fourier transform and the inverse transformation thereof to obtain a residual disturbance gravitation value; calculating a low-order model disturbance gravitation and a final disturbance gravitation to obtain the final disturbance gravitation; comparing a disturbance gravitation precision with a calculation efficiency, at the same time, comparing the efficiency difference with a conventional method. The method of the present invention is easy to operate and apply, can effectively solve the problem of obtaining the disturbance gravitation of earth's any height level rapidly and at the high precision, facilitates the rapid and real-time positioning and control of various spacecrafts when the spacecrafts move, guarantees the safety of the spacecrafts, and possesses a very strong use value.

The present invention provides a binary convolutional device and a corresponding binary convolutional neural network processor. The binary convolutional device comprises: an XNOR gate taking elements in an employed convolution kernel and corresponding elements in data to be subjected to convolution as input, wherein the elements in the employed convolution kernel and the corresponding elements in the data to be subjected to convolution are both binary forms; and an accumulation device taking the output of the XNOR gate as input and configured to perform accumulation of the output of the XNOR gate to output a binary convolutional result. According to the technical scheme, the bit wide of data for calculation is reduced in the operation process so as to reach the effect of improving operation efficiency and reduce storage capacity and energy consumption.

The invention provides a target classification method based on a convolutional neural network. the target image is directly traversed without adopting a convolution kernel; instead, a sliding window with the same size as the output feature image is adopted to traverse the whole target image row by row and column by column; pixel points corresponding to the target image are extracted to serve as sub-images; correspondingly each characteristic parameter of the convolution kernel is multiplied by each sub-image to obtain an intermediate image; and finally, the sum value of the intermediate imagesis taken as an output feature image. Under the premise that a convolution result which is the same as that of an existing convolution implementation mode is obtained, convolution operation is dividedinto multiplication and addition operation of a single point, the number of times of address hopping when a microprocessor reads data in the convolution implementation process can be reduced to the maximum extent, and then the hardware processing efficiency is greatly improved.