79 results about "Kernel size" patented technology

One embodiment provides a compute apparatus to perform machine learning operations, the compute apparatus comprising a hardware accelerator including a compute unit to perform a Winograd convolution, the compute unit configurable to perform the Winograd convolution for a first kernel size using a transform associated with a second kernel size.

An image descreening process first smoothes the image, where smoothing is accomplished by applying a convolution with a low pass filter (LPF) kernel, which is a parameter to the descreening function. Using the smoothed image, a determination is made for each pixel for which pixels around it participate in the modified filter. For a current pixel, a window is considered having the size of the LPF kernel, with the current pixel at the center. A threshold T1 which is given as a parameter, is used to mark the pixels in the current window. Considering a pixel in the window, if for all color components the difference between this pixel value to the center pixel value is less than T1 in absolute value the pixel is marked with a 1. Otherwise, the pixel is marked with a 0. Finally, an adaptive version of the LPF is applied. If the number of pixels marked with a 1 in the window is less than a third of the kernel size, the original pixel value is restored. Additionally, for a color component for which there is a small change in values within the original (non-smoothed) window (i.e. the difference between the maximal value to the minimal value in this component is less than another threshold T2), the value of this color component is restored. If these conditions do not hold, a new value for each component is determined. To be the convolution of the original window, the LPF kernel is masked with the 0/1 markings from the second step. That is, the modified convolution uses an adaptive kernel which is identical to the LPF kernel in the locations marked with one, but has zero entries in the locations marked with zero.

A lane recognition image processing apparatus can improve lane marking recognition performance by preventing false detection by addition of a condition without changing the basic principle of an one-dimensional image filter. A search area is set for each lane marking with respect to images stored in a image storage part through a window. A candidate point extraction part extracts candidate points of each lane marking from the search area thus set. A lane marking mathematical model equation is derived by approximating sets of extracted candidate points by a mathematical model equation. The candidate point extraction part includes a kernel size setting part, a filtering part that outputs, as a filtering result, the smaller of differences between the gray value of a pixel of interest and those of pixels forwardly and rearwardly apart a kernel size from the pixel of interest in a scanning direction, respectively, and a binarization part that binarizes the filtering result.

The invention discloses a super-resolution reconstruction method based on fusion of a multi-level feature map, and the method comprises the following steps of employing the idea of a dense network toconstruct a feature extraction network for generating the multi-level feature map; performing dimensionality reduction on the connected feature map by using a convolutional neural network with a convolution kernel size of 1 * 1, fusing the feature map, performing feature extraction of the network on the basis to obtain a fused multi-level feature map, and using a sub-pixel convolutional neural network as an upsampling operator to obtain a high-resolution reconstructed image. In the training process, a perception loss function is used as a minimization target to generate a high-resolution imagemore conforming to visual perception. According to the method, the defect that an existing super-resolution reconstruction algorithm cannot fully utilize a multi-level feature map is overcome, localand overall information in a low-resolution image obtained by a feature extraction network can be fully utilized, and a high-resolution image can be accurately and quickly reconstructed from the low-resolution image.

A magnetic resonance method and system are provided for providing improved multi-band (MB) magnetic resonance imaging. The adaptive MB imaging can be achieved by providing one or more modified multi-band excitation pulse sequences that include at least either one nullified “dummy” slice within a slab that is not excited simultaneously with the other slices during a single multislice acquisition sequence, or one excitation slice group that utilizes a non-uniform slice spacing between simultaneously excited slices. Adaptive GRAPPA or slice-GRAPPA kernel sizes can also be used during image reconstruction to improve speed without excessive point spread blurring or MB reconstruction failure. A total leakage factor (TLF) can also be determined based on test images using modified MB excitation sequences, and used to improve the adaptive MB procedure.

A multi-dimensional image data enhancement system is configured to reduce the effect of blanking areas on the large kernel processing of pixel data. Thus, when pixels comprising blanking data are processed, additional blanking data can be added to the blanking data such that the additional “fictional blanking data” will occupy an area within the kernel size instead of image data in an adjacent frame or non-related areas of the current frame.

A face feature extraction method based on depth learning comprises the following steps: 1, initializing an input image with the dimension as W*W, a convolution kernel size as K*K, a step length as S,a filling size as P, an output number as NUM, and weight in each layer as [theta] i; 2, preparing face image data; 3, building a deep learning face feature extraction network with 46 layers; 4, inputting the prepared face image data into the face feature extraction network formed in step 3, and training a softmax classifier; 5, starting the face feature extraction network, carrying out network training for a time of T, using fine tuning skills to improve the precision of the face feature extraction network, finally obtaining a weight system [theta] as the solved model; using the model to extract features from known samples, and finishing the flow.

The invention discloses a transmission line anti-outer damage vehicle identification and danger behavior discrimination system and method; the system comprises a front end power transmission line intelligent anti-outer damage early warning device, a rear end center management platform and a mobile client end; the front end and rear end of the system are respectively provided with a depth learningmodule; the depth learning modules in front and rear ends can operate in synergy, thus finishing transmission line anti-outer damage vehicle identification and vehicle danger behavior early warning and discrimination; the model and parameters of the depth learning module in the front end are provided online through rear end training; the center management platform determines the layers and convolution kernel size of the depth learning network, and trains the depth learning network. The invention also provides a discrimination method of the system; the system and method can prevent training costs of the front end depth learning network, can reduce the complexity of the front end system, and can solve the problems that the rear end intelligent analysis is high in video transmission cost andlow in timeliness.

A high speed image processing apparatus is created through the use of cascaded elongated filters. The processing speed of the filters is kernel size insensitive, enabling use of general purpose computing facilities to process high resolution, monochrome, and multi-spectrum images. Elongated filters described include both linear and non-linear filters. Very large kernel and multi-dimensional image processing is accomplished with reduced complexity and portable programming instructions.

The invention discloses a method for improving video resolution through a convolutional neural network. According to the method, a video super-resolution reconstruction model based on the convolutional neural network is constructed according to a video image sequence, and in the model construction process, super parameters, including convolution kernel size, the number of neural network layers, etc. of the convolutional neural network are set based on the characteristics of the video image sequence; and then a single-image super-resolution method is utilized to generate image training weightsused for initializing weight parameters of the video model, redundant information among video frames is fully utilized, multiple frame video images are used as input of the convolutional neural network model, and finally a high-resolution video is obtained through an incremental iteration method. Through the method, the video super-resolution model is high in training speed and high in predictionprecision; and a comparison test verification result indicates that compared with other super-resolution algorithms, the peak signal-to-noise ratio and structural similarity of the images reconstructed through the method have comprehensive optimal results.

The present invention provides a hyper-parameter determination method for the critical convolutional layer of a remote-sensing classification convolution neural network. The method comprises the steps of constructing a convolutional neural network sample set; constructing a convolutional neural network structure; deterring the hyper-parameters of the critical layer of the convolutional neural network; selecting one convolutional layer as the critical layer, presetting the convolutional kernel size of the critical layer, and calculating the convolution scale; based on the convolutional kernel of the critical layer and the convolution scale, calculating the convolution step length according to a preset rule; presetting the convolutional kernels of other convolutional layers to be kernel size, and presetting the convolution step lengths of other convolutional layers to be 1; and conducting the mean-value down-sampling or the maximum-value down-sampling as the subsequent down-sampling. According to the technical scheme of the invention, based on the image input size and the convolution kernel size, the convolution scale concept is proposed and is adaptive to the remote-sensing spatial scale. On the above basis, an input size and convolution scale-based method for jointly determining the hyper-parameters of the critical layers is provided. In this way, the parameter adjustment time required for the algorithm is reduced, and the object-oriented remote-sensing classification precision is improved.

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 relates to a TCN-based wind power prediction error interval evaluation method, which comprises the steps of obtaining and normalizing wind power original data, and dividing the wind power original data into a training set and a test set; initializing the weight of the TCN network, and setting the number of iterations, the number of residual modules, the number of network layers, an expansion coefficient, a convolution kernel size, a learning rate and parameters of the number of hidden layer neurons; inputting the training set data into the TCN in batches, calculating an output error of an effective historical length, and performing back propagation on the error to update TCN parameters; completing the training until the TCN network precision meets the artificially given erroror reaches the number of iterations; and inputting test data and outputting a prediction result to realize real-time error prediction of the wind power. Error prediction can be realized by means of the flexible receptive field and learning capability of the TCN, thereby providing a basis for real-time scheduling.

A method and system for generating shadows for a graphics processing unit. Specifically, the method determines whether a potential blocker occludes light from reaching a point of a scene in an image space. The light is generated from a light source. A width of a corresponding penumbra is determined for the point. The width is based on a width of the light source, a depth of the potential blocker from the light source, and a depth of a receiver from the light source. The receiver includes the point. A percentage closer filtering kernel size is scaled in proportion to the width of the corresponding penumbra. Thereafter, percentage closer filtering is performed for the point using the kernel size that is scaled in order to shade a pixel corresponding to the point.

The invention discloses a construction method of a new deep learning model, in particular to a deep learning model comprising two CNNs with different structures and converging quickly. The convolutionlayers of the two CNNs in the deep learning model constructed by the invention, convolution kernel size, The number of pooling layers and the full connection mode are different, and each CNN shares the feature information obtained from its learning once. Before sharing or accepting the feature information, the two CNNs undergo a batch normalization process. If the number of channels of the two CNNs is different when the feature information is shared, the number of channels is adjusted, and then the batch normalization process is performed. The current research direction of improving performance of deep learning model is to deepen the network deep of the model, The deep learning model provided by the invention improves the performance of the model, greatly accelerates the convergence speedof the network, reduces the parameters of the model, and reduces the serious dependence of the excellent deep learning model on the computer performance without constructing a deeper network.

The invention provides an automatic convolution kernel size determining convolutional neural network-based hyperspectral image classification method. The method comprises the following steps of: firstly clustering a plurality of image block samples with different sizes by using a K-means algorithm; designing a new evaluation index for a clustering result of the samples, the sizes of which are different and category quantities of which are uniformly distributed, and evaluating the clustering the result of the samples with different sizes by using the evaluation index, and determining a self-adaptive convolution kernel according to the evaluation result; and finally training a convolutional neural network of the self-adaptive convolution kernel, and classifying hyperspectral images by utilizing the trained network. Self-adaptive convolution kernels obtained by clustering and index evaluation are capable of representing data information more effectively, so that better hyperspectral imageclassification results can be obtained by utilizing the method.

The related file system of an OS comprises: an interface conversion unit on entrance, a management unit, a physical file system unit, a top drive unit, and an interface conversion unit on exit; all connected in turn. This invention reduces VFS unit in Linux standard file system and kernel size, can be used to process different custom special affairs, and realizes cross-OS transplant.

The invention discloses a dam deformation prediction system and method based on a convolutional neural network. According to the convolutional neural network for dam deformation prediction, the network parallel computing capacity is fully used, two-dimensional data is input into multiple paths in parallel, multiple convolutional layers and pooling layers are alternately cascaded in each path, andin order to obtain more characteristics, different convolutional kernel sizes can be set; a simple path only cascading a convolution layer and a pooling layer is added to extract features different from other paths, and meanwhile, the features extracted by the paths are not necessarily complementary, so that additive operation is used in the system when output results of the paths are combined. According to the method, the automation level of dam deformation prediction is improved, and the efficiency and speed of deformation prediction are improved.

The invention provides a method for determining the convolution neural network convolution kernel quantity based edge detection. The method comprises steps that the convolution neural network layer quantity and convolution kernel sizes are set; after an image dimension is reduced to 30*30, and edge detection for a training image is carried out to acquire an edge image; edge block extraction and statistics analysis processing on the edge image is carried out according to the convolution kernel size of each convolution layer to acquire the convolution kernel quantity of each layer; lastly, an integral convolution neural network is constructed to train an RGB-D data set. The method is advantaged in that in a convolution neural network construction process, the convolution kernel quantity of each convolution layer can be automatically determined, efficiency of designing the convolution neural network is improved, adaptive capability of the convolution neural network is enhanced, and the designed convolution neural network is made to be able to adapt to different-size data sets.