195 results about "Convolution filter" patented technology

A technique for controlling a loudspeaker system with an artificial neural network includes filtering, with a deconvolution filter, a measured system response of a loudspeaker and a reverberant environment in which the loudspeaker is disposed to generate a filtered response, wherein the measured system response corresponds to an audio input signal applied to the loudspeaker while the loudspeaker is disposed in the reverberant environment. The techniques further include generating, via a neural network model, an initial neural network output based on the audio input signal, comparing the initial neural network output to the filtered response to determine an error value, and generating, via the neural network model, an updated neural network output based on the audio input signal and the error value.

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.

This invention provides self-clocking glyph shape codes for encoding digital data in the shapes of glyphs that are suitable for printing on hardcopy recording media. Advantageously, the glyphs are selected so that they tend not to degrade into each other when they are degraded and/or distorted as a result, for example, of being photocopied, transmitted via facsimile, and/or scanned-in to an electronic document processing system. Moreover, for at least some applications, the glyphs desirably are composed of printed pixel patterns containing nearly the same number of ON pixels and nearly the same number of OFF pixels, such that the code that is rendered by printing such glyphs on substantially uniformly spaced centers appears to have a generally uniform texture. In the case of codes printed at higher spatial densities, this texture is likely to be perceived as a generally uniform gray tone. Binary image processing and convolution filtering techniques for decoding such codes also are disclosed, but this application focuses on the codes.

The invention discloses an elimination method of slurry pulse signal pump stroke noise. The order formula (N=Fs/Fo) of a comb filter is utilized, wherein Fs is the sampling frequency of a slurry signal, and Fo is a product of the movement rate of a single slurry pump piston and the number of slurry pumps, and the order of the comb filter is determined; when the pump stroke noise is eliminated, a filtering process of a slurry pulse signal is realized by utilizing MATLAB software, and comprises the following steps: 1) collecting the slurry pulse signal on site; 2) according to the signal to noise ratio of the signal, determining the bandwidth (BW) of the comb filter, designing the required comb filter according to the BW and the determined order; 3) converting the slurry pulse signal to enable a signal value to be within an interval (-1,1), and obtaining a signal X(n); 4) loading the signal X(n) into the comb filter to carry out convolution filtering to obtain a filtered signal Y(n); and 5) plotting, and observing a signal filtering situation. The invention breaks through the restriction of a hardware circuit, the MATLAB software is utilized to effectively realize the filtering process of the slurry pulse signal, and meanwhile, the comb filter is applied to the slurry pulse signal so as to further enhance a filtering effect.

The invention relates to a deep convolutional neural network method based on information lossless pooling. The method is used for image classification, and comprises the steps of collecting different kinds of images and marking image categories as image label information; performing image set division, and dividing the collected images into a training set, a verification set and a test set; designing a convolutional neural network structure based on information lossless pooling, including the number of convolutional layers and the number of information lossless pooling layers that are used, designing the number of filters in the convolutional layers, designing Gaussian smoothing filter parameters of the information lossless pooling layers, the pooling window size and a convolutional filter structure for feature fusion, designing the number of network training loop iterations and network final convergence conditions, and initializing network parameters; and inputting training data to the network in batches for calculation and training.

A system and method for an improved convolutional layer in convolutional neural networks is provided. The convolution is performed via a transformation that includes relocating input, relocating convolution filters and performing an aggregate matrix multiply.

A computing device for detecting an object of interest in an image is provided. The computing device includes a memory device and a processor to process image data by evaluating pixel contrast to facilitate detecting an edge transition in an image, by identifying neighboring pixels that are above a predetermined contrast value, which represent a potential object of interest, filtering the pixels based on color information to identify areas where there is a color edge transition, applying a convolution filter mask to yield regions of pixels having a localized concentration of contrast and color changes, applying a threshold filter for identifying those pixel regions having values that are greater than a predetermined threshold value and grouping the identified pixel regions into one or more groups representing an object of interest, and generating an alert to notify an operator that the object of interest in the image has been detected.

Disclosed are a virtual stereo synthesis method and device. The method comprises: acquiring at least one voice input signal at one side and at least one voice input signal at the other side; respectively conducting ratio processing on a left ear component of a preset head related transfer function (HRTF) and a right ear component of the preset head related transfer function (HRTF) of each of the voice input signals at the other side, so as to obtain a filter function of each of the voice input signals at the other side; respectively conducting convolution filtering on each of the voice input signals at the other side and the filter function of each of the voice input signals at the other side to obtain filtering signals at the other side; and synthesizing all the voice input signals at one side and all the filtering signals at the other side to form virtual stereo signals. By means of the above-mentioned method, the present application can improve the tonal coloration effect and reduce the calculation complexity.

A defect detection method comprising, irradiating light from a light source in an optical system and obtaining a plurality of optical images of a sample having a repeated pattern of a size smaller than a resolution of the optical system; while changing the conditions of the optical system, performing correction processing for the optical images with the use of at least one of a noise filter and a convolution filter; shifting a position of the other optical images based on any of the plurality of optical images, obtaining a relationship between shift amounts of the other optical images and a change of correlation of a gray scale value between the plurality of optical images, and performing positional alignment of the optical images based on the shift amount obtained when the correlation is highest, performing defect detection of the sample with the use of the optical images after the positional alignment.

The invention discloses an incomplete projection CT image reconstruction method based on VAEGAN. The method comprises the following steps: firstly, a sinusoidal image of complete projection data is used for training a VAEGAN model to obtain a VAEGAN model capable of generating a high-quality sinusoidal image; secondly, using the trained model to predict the missing part of sinusoidal graph of incomplete projection data, the completed projection data can be obtained. Finally, the CT image is reconstructed from the completed projection data using the convolution filtered back projection (FBP) method. The invention can predict the missing projection data and further reconstruct the CT image with high quality in accordance with clinical diagnosis.

Imaging apparatus (26) is provided for use with an image sensor (24). The apparatus includes a non-linear mapping circuit (42), which is configured to receive a raw stream of input pixel values generated by the image sensor and to perform a non-linear mapping of the input pixel values. to generate a mapped stream of mapped pixel values, and a linear convolution filter (44), which is arranged to filter the mapped stream of mapped pixel values to generate a filtered stream of filtered pixel values. Other embodiments are also described.

A finger vein image recognition method based on fusion of local coding and CNN model is proposed. The method comprises the following steps: extracting the finger vein image from the finger vein image;using the encoder based on weighted symmetric local graph structure to encode the image: the reconstructed encoded convolutional filter is obtained; establishing an improved convolution neural network mode; performing feature extraction of a digital vein ROI image; measuring the similarity of the ROI images of digital veins to be matched. The finger vein image recognition method integrating the local coding and the CNN model provided by the invention can not only solve the problem of variable finger posture to a certain extent; experimental results on two finger vein databases show that the proposed method is feasible to some extent, which not only has good matching effect, but also reduces the number of parameters that can be learned, reduces the learning cost and improves the matching efficiency.

The invention relates to a deep neural network method based on space fusion pooling. The method, which is used in image classification, includes the following steps: collecting images of various categories, and marking the category of each image as image tag information; carrying out image set division: dividing the collected images into a training set, a verification set and a test set, wherein the training set is used to train a convolutional neural network; designing a deep neural network structure for image classification (including the number of convolution layers and the number of space fusion pooling layers used), designing the number of filters in the convolution layers, designing the fusion function form, space sliding step length, pooling function, pooling window size and step length in the space fusion pooling layers, designing a convolutional filter structure for feature fusion, designing the number of network training loop iterations and the network convergence conditions, and initializing the network parameters; and inputting training data to the network in batches for calculating and training.

A technique for an equalizer filter with dynamically configurable convolution filter is described. The input to a transmitter chain is modified by an equalizer filter, prior to being applied to the transmitter. The equalizer filter modifies and smoothen the amplitude of the signal. The modified and smoothen signal has its peaks reduced which results in lower Crest Factor. The input to the reconditioning equalizer filter could be a baseband, an intermediate frequency (IF) or radio frequency (RF) signal. In the case of an IF or RF signal, it needs to be down converted to baseband before applied to the equalizer filter.

The invention discloses a test paper surface texture defect detection method based on gray scale gradient clustering, which comprises the following steps: collecting a frame of image of test paper, and carrying out graying and median filtering preprocessing on the image; performing binary segmentation on the image based on a four-point gray scale dynamic threshold, and extracting a test paper areathrough a difference method; carrying out Gama gray scale enhancement on the image and filtering out part of periodic textures by adopting a Gaussian low-pass filter; constructing a single-directionGaussian kernel function to carry out convolution filtering on the image in the vertical direction; calculating the gradient grad _ x of the image in the horizontal direction; dividing the test paperarea into n columns of subareas along the horizontal direction, and calculating the position of a gradient maximum value area of each subarea; carrying out clustering calculation on the position of the gradient maximum value area of each sub-area in the vertical direction, and marking the area with the area clustering number reaching the threshold range as a texture defect area; and judging whether the test paper is qualified according to the marked area. The method has the advantages of high detection speed, high detection precision, good robustness and the like.

The invention discloses a fabric defect detection method based on a single-classification support vector machine (SVM). The method comprises the following steps: obtaining a defect-free fabric image, optimizing parameters of a Gabor filter by use of an RDPSO algorithm, and constructing a single optimal Gabor filter most adapted to a defect-free fabric image texture feature; optimizing parameters of the single-classification SVM by use of the RDPSO algorithm; performing Gabor convolution filtering on a fabric image to be detected; extracting one group of texture features on the image after filtering on the basis of GLCM; and performing defect determining by use of the single-classification SVM. According to the invention, by use of the single optimal Gabor filter, the detection speed can be effectively improved and the requirement for real-time performance of a system is ensured; and by taking the single-classification SVM as a defect determining method, the problems of local extreme values, over-learning, under-learning and the like by use of a conventional statistical mode identification method can be avoided, the generalization capability of a system can be effectively improved, and the requirement for detection accuracy of the system can be guaranteed.

The invention discloses a deep convolutional neural network method based on deterministic discarding connection. The method comprises the following steps: step one, collecting a training data set and carrying out tag marking on the data; step two, setting a structure of a convolutional neural network; step three, determining a threshold t that needs to be discarded by a convolution filter weight, initializing the convolution filter weight and determining an iteration frequency; step four, carrying out convolution operation to obtain a feature maps of all convolution layers and inputting the feature map of the last convolution layer into a classifier to obtain a data classification result so as to complete forward calculation; step five, carrying out updating and convolution filter weight optimization continuously by using a back propagation algorithm from the last convolution layer to the first convolution layer to reduce a training error; and step six, repeating the steps four and five and carrying out iteration to optimize parameters of the convolutional neural network continuously till meeting of a terminal condition.

The invention discloses a convolution neural network learning algorithm based on a limit learning machine, the algorithm is based on the idea of self-encoding to learn the convolutional filter with bias. Firstly, the data matrix is used to generate the standardized training data with the mean value of 0 and variance of 1. Secondly, the data matrix is used to generate the standardized training datawith the mean value of 0 and variance of 1. Then the convolution deviation is processed, the intercept term is added, and the normalized input and intercept terms are reconstructed so that the objective matrix becomes a formula shown in the specification. Given input and objective matrix, and reshaping matrix are used to obtain a filter. The invention is based on an automatic coding limit learning machine, which learns a convolution filter, and is used for training arbitrary convolution neural network to work, dealing with the deviation of the filter, and reconstructing a standardized input with an intercept term. The invention is a hierarchical training process, does not need the entire classification model to extract arbitrary features, improves the training speed, realizes a competitive result in the generalization performance, and exceeds the BP in the training speed. CNN; while memory consumption is reduced.