KR20180123810A - Data enrichment processing technology and method for decoding x-ray medical image - Google Patents

Abstract

Various types of data are exponentially increasing through an IoT, a cloud, and the like which are performed through a wired/wireless network and appears in various industrial fields such as factory automation, medical equipment, and the like. The data can be used in image recognition, data analysis, and the like through an artificial intelligence function to be effectively used in interconnected information. The present invention relates to a computer structure in which a deep neural network circuit and a processor are coupled through a machine learning technique through abstraction through the combination of several nonlinear conversion techniques, and an image conversion and analysis processing method using the same. According to the present invention, data and image data of an image can be analyzed and processed by extracting various information collected through a cloud server and information of a feature point, a color, a line, a shape, and the like of structured data of various information, images, and the like corrected through a cloud server, converting an image of an object into 2D coordinates, and repeatedly performing processes of structuration, filtration, nonlinearization, extraction on software that includes a neural network circuit outputting the coordinates into a packet format and a plurality of processing units and performs multiple recognition circuit operations between pixels of an image patch formed based on the extracted feature point and multilayer-based data.

Description

TECHNICAL FIELD [0001] The present invention relates to a data processing method and a data processing method for X-

The present invention relates to a computing device, and more particularly to a client device with a neural network and a system containing it.

Recently, computers provide machine vision or object recognition. Object recognition provides the user with various useful tools. For example, object recognition relies on an algorithm that includes a neural network. That is, the user device can recognize the object included in the input image using the neural network. Generally, a neural network trains to recognize objects using training images. This object recognition process can have a higher discrimination power as more training images are used.

In general, neural networks include systems of interconnected " neurons ". Neural networks can learn the machine as well as calculate values from inputs and pattern recognition as a result of their adaptive nature.

Neural networks for image recognition require data processes for learning and recognition that require memory and processing intensive. Therefore, a lot of calculations are needed. In fact, the weight values require memory for storage and processing, while performing computational processes.

As can be expected, the performance of the neural network improves as the size of the training data set increases. Unfortunately, mobile devices such as smart phones have limited memory and processing capacity. Thus, increasingly popular mobile devices generally can not use image recognition technology. Accordingly, there is a need for a method of improving the performance of a neural network in a computing device having limited resources.

It is an object of the present invention to provide an image analysis and transformation apparatus accompanied by a neural network capable of performing object recognition through limited resources, and a system including the same.

According to an aspect of the present invention, there is provided a client apparatus including a neural network, the apparatus including a processor, a memory for storing the neural network received from the server system, and an input device for receiving the neural network, The neural network is trained for the client device in the server system.

In an embodiment, the input device captures an image and stores the image input data in the memory.

As an embodiment, it further includes a multilayer perceptron classifier for mapping the image input data.

The learning data storage device using an image includes an input unit for receiving a plurality of images including an image, a control unit for building a storage based on the received images, and a database for storing object-classified data, A classification unit for grouping the detected images and tagging the grouped images, a classification unit for classifying the tagged images according to dates, And a DB constructing unit for receiving the multimedia data corresponding to the age of the sorted images from the database and grouping them.

As an embodiment, the learning method of the neural network includes one of inverse delivery and automatic decoding.

According to an aspect of the present invention, there is provided a server system including an input device for receiving a training image, at least two layer pairs, each layer pair including a convolutional layer and a sub-sampling layer, Wherein the neural network performs a quantization of a temporary weight at the convolutional layer and generates a temporary feature map at the subsampling layer in response to an input applied to the convolutional layer, Performs quantization of the weights in a multi-layer perceptron classifier, and generates a classifier output in response to the temporary feature map applied to the quantized weight multi-layer perceptron in the multi-layer perceptron classifier.

According to an embodiment of the present invention, a client apparatus including a neural network capable of performing object recognition through limited resources and a system including the client apparatus can be provided.

1 is a block diagram illustrating a neural network according to an embodiment of the invention.
2 is an exemplary diagram illustrating a data structure of a sample convolutional neural network used to perform a simulation according to the present invention.

The types of neural networks include having one or two layers of unidirectional logic, having complex multiple inputs, having feedback loops in many directions, and having many layers. Generally, such systems use programmed algorithms to determine the control and organization of their functions. Most systems use " weights " (which can be expressed as values) to change the throughput parameters and the various connections to the neurons. Neural networks can learn spontaneously from previous training completed through the use of sets of training data.

This convolution operation has a layer of convolution units, so that a contour line can be determined from a combination of contour lines, a figure can be found from a combination of contour lines, and an object can be found step by step. In a convolution neural network, a pooling operation finds out which local pattern exists in an area from data such as an image, and repeats it to find a characteristic of the data.

101: Sensor node
102: Data preprocessing
103: Analysis of search data
104: Machine learning
105: Complex Event Processing
106: Report
107: Neurong Network
301: Input image
402: convolution layer
403: SBSA layer
406: Result image
407: Neural network

Claims (9)

Extracting feature vectors from the data and image frames;
(DNN) model by setting feature vectors as inputs of the deep neural network, outputting the feature vectors extracted from the target frame, and training the weights of the deep neural networks based on each feature vector set to the input and output And analyzing the learned data based on the generated deep neural network (DNN) model. The method according to claim 1,
Wherein the input device captures an image and stores image input data in the memory. The method according to claim 1,
Image transformation software further comprising a multilayer perceptron classifier for mapping image input data. The method according to claim 1,
Wherein the neural network comprises a convolutional neural network. The method according to claim 1,
Wherein the neural network generates a feature map. 6. The method of claim 5,
Wherein the feature map comprises a plurality of weight values derived from an input image. Wherein,
Estimating a phase of a lost packet and a log power spectra based on the deep neural network; And restoring the lost packet by inverse Fourier transforming the estimated phase and log power spectra. The parallel processing processor allocates to each processing unit an image area in which eight images of the object are vertically divided into eight processing units, and the pixels in each line of the image of the object process the image at the same time, A computer architecture that combines a neural network circuit with a parallel processing processor. Assigning to each processing unit an image area in which the entire image is vertically 8-divided for the plurality of processing units, and wherein pixels in each line of the image process the image simultaneously to output or update the image line by line, A method of object recognition processing using a computer structure combining a parallel processor and a parallel processor.

Images