KR20090035672A - Remote control using prediction neural network image processing - Google Patents

Abstract

An instrument controller using a shape example and an image pattern learn is provided to reduce the load of a processor and reduce a processing time by analyzing the generation of the pattern based on the pattern. An inputting unit inputs and image analyzes the image. A motion analyzing unit extracts a motion of the analyzed image and predicts the motion. A data unit outputs a control command as data. A controller controls each configuration part.

Description

Remote Control using Prediction Neural Network Image Processing}

The present invention is a technique in the field of image processing and artificial intelligence used in the control of the device to analyze the user's motion continuously changed by recognizing the changing motion pattern of the user in the technology for controlling TV or home appliances.

Methods of controlling the device can be divided into wired control and wireless control. In general wired control, the device and the controller use a data communication line to control physical electrical and communication signals. In the case of control using wireless, there is a communication method using RF and IR. This control is a technology that eventually transmits the promised control signal to the device and uses it for the control of the device.

This control technology is also implemented as a biometric technology by the development of IT technology. A typical example is voice recognition technology. Speech recognition technology is one of the recognition technologies that give commands, not identify people, mainly in devices such as mobile phones or controlled areas. However, these control techniques are very vulnerable to noise such as white noise, single sound or ringing sound when applied to industrialization.

The control method of the device less affected by the environment than the voice recognition is a control technology using the image. Control of a device using an image is to analyze the image and use it for device control. Methods of analyzing images include gesture recognition for controlling a device according to a specific behavior pattern, and shape recognition technology for recognizing a specific shape.

The present invention recognizes the pattern of motion more intelligently by predicting the change value of the image itself instead of predicting the movement of the point by constructing the neural network of the predicted image to be changed next in the reference image rather than analyzing the motion during the process of recognizing a specific shape. It is an artificial intelligence technology used for device control by calculating the change value.

Gesture recognition technology recognizes a sequence of motion sequence patterns when describing a specific action or action as an object or part of the body. Such gesture recognition technology is basically a technique of specifying a singularity of an object through an image or a motion sensor to measure the variation of the singularity and detecting the characteristic according to a movement pattern or a shape pattern.

In the field of recognizing a specific gesture, there are a motion sensor that uses a sensor to make data and a technology of tracking an object by using an image of the object itself or by attaching an object that emits a specific frequency.

Recognition of objects by analyzing images can be divided into neural network engines that use general pattern analysis and image processing techniques, and compares specific patterns of given data with previously learned patterns.

When we recognize a certain gesture, we analyze the motions between still images, but we cannot extract the pattern of continuous images. In other words, when we analyze the movements of hands through the images, we measure the change, that is, the current image with the background image. To analyze a particular shape. This gesture recognition basically assumes that the input image is a specific object. However, it is very important to distinguish specific objects for real life applications, for example, a technique of tracking a finger with a mark that emits a frequency of a specific area can be used. However, this pattern is not performed as prescribed.

Therefore, similar patterns other than these patterns should also be recognized.

Techniques for analyzing motion and dataizing it typically use equipment called motion capture. Motion capture is a device that analyzes the movement of an object and analyzes it by converting it into numerical data. The motion capture device oscillates and detects the frequency of a specific area of a sensor or instrument to extract motion, but it can only calculate the current movement and cannot analyze a specific pattern. When inferring a specific pattern and predicting it, the analysis time is shortened by predicting the motion added after the current motion with the predetermined learning data. The shape prediction neural network engine can be used to predict certain patterns.

Existing image analysis or gesture control technology has one fatal drawback: it can analyze the current motion and apply it to a predetermined pattern, but it cannot be calculated by inferring the motion in the next time series. If you infer the next pattern by intelligently analyzing the time series pattern, that is, the pattern of change in motion, you can shorten the processing time and improve the accuracy by reducing the cost of product, miniaturization and processor resources by reducing the processing time. .

Literature Information of the Prior Art

[Document 1]

Audiovisual-based User Recognition and Gesture Recognition Technology

                                               Daedeok Special Zone Technology

[Document 2]

A Study on Vehicle Tracking for Intelligent Transportation System

                                       Department of Multimedia Information, Sungduk University

[Document 3]

New Iris Control Method and Hardware Implementation for Image Detector

                                      Korean Information Science Society 28th Academic Paper

[Document 4]

J. Malik, S. Russell, J. Webwe, T. Huang, D. Koller "A machine Vision

           Based Surveillance System for California Road ", PATH

           project MOU-83 Final Report

In a method of controlling a device through image analysis, the control technology of the device recognizes a promised gesture of a specific object and calculates data based on the predetermined gesture. The problem is that processing time must be real-time in recognizing specific movements. The most time-consuming part of the processing algorithm that analyzes the image and finds an object to determine if there is a predetermined pattern of the current image is to find a specific object. The correlation between the current image and the next image cannot be taken into account, and if we can predict this link relationship and the prediction of the next pattern first, we can process it more accurately in real time.

In the technology for controlling a device through an image according to the present invention, the next pattern occurrence is inferred based on a previously learned pattern to shorten the processing time and reduce the load of the processor so that the next motion of the image is a motion vector of the object. It is a technique that can be predicted through shapes.

Among the general image processing techniques, many predictive algorithms, that is, prediction methods by analyzing a moving direction, have been published. Predicting the motion vector through time series analysis between the simple motion vector and the frame, but the point analysis in such an image is limited in reducing the processor's excessive processing time in interframe processing. Vector analysis also goes through basic preprocessing.

Therefore, if the next shape is predicted based on the shape without a simple vector operation and a simple matching is performed, the processing time required for the vector operation can be reduced. Each shape uses the predetermined shape itself as an input variable in the reference image. [Fig. 1] The primary input variable in the neural network is defined as the next predicted reference image immediately after the previously learned reference image. Can be inferred quickly.

It is possible to obtain a drastic reduction in processing time and accurate data when inferring a specific behavior pattern itself in a frame rather than processing a general motion vector as a basic data and processing the block by following a differential motion. will be.

Let's look at the simplest derivative among the methods of extracting features from an image.

y '= lim (Δx → 0) Δy / Δx = lim (Δx → 0) {(x + Δx) ＾ n-1 + (x + Δx) ＾ n-2x + .... + x ＾ n-1}

When the total repetition time is included in the vector image as described above, when the simple difference image is X, the temporal increase rate of the vector operation is X + Xn + Xm, which means that it is a waste of time more than three times.

Real-time processing is very important when predicting the movement of an object through the image and using it for the control of the device. When analyzing the motion pattern of the object in a continuous next pattern as in the present invention, the temporal deviation can be drastically reduced. It makes it possible to reduce the weight and real-time processing of the processor, and to realize the general purpose economically. Therefore, it is possible to realize the control of all electronic products that use the device control as the image at low cost.

The hardware configuration is as follows [Figure 3] is a hardware configuration that can acquire and control the general motion. Input device for receiving images and analyzing them [Fig. 3] A motion analyzer extracting and predicting motion from the analyzed image [Fig. 3b], a controller for controlling them [Fig. It consists of a data section [Fig. 3].

In the present invention, the hardware configuration for analyzing a specific image can be divided into an image input unit, a controller, an analyzer, a data output unit, and the like.

The configuration diagram of the processor is as follows.

2A is a processor for preprocessing an image. 2b is an input image and FIG. 2b is an extractor for extracting motion. [Figure 2d] is an extractor that can extract images from the current frame. 2 is motion pattern data previously learned for the prediction data. 2 is data calculated by leaking the pre-motion of the next frame using the predicted pattern.

Usually, the motion vector of a specific block is used to detect motion. The motion vector is a weight of a vector moving in a next image based on a point based on a distribution of a specific point between frames.

In Fig. 4, N1 to N4 represent the moving vector values of a certain point, and the vector of moving points is projected on one image in the thin continuous frame of [Fig. 4A]. The motion vector of these objects analyzes the points and when analyzing the images, these points must be collected and recombined into a larger image. However, in the general motion vector equation, the processing time takes about three times the size of the entire image, and the shape tracking is accurate, not the vector, when predicting the motion itself rather than image mobility. Shape tracking is to reduce the time to analyze the shape of the current image by inferring a motion pattern when recomposing the segmented image from the next input image by refining it by specifying a specific numerical shape in the current frame. .

As shown in FIG. 5, the extraction of the finger direction of the extracted object [FIG. 5] is shown, not the coordinate of a simple point when the object is predicted. The next shape, that is, the data can be calculated faster by predicting N3, which calculates the moving value based on N1 to N2.

1

1 is a general neural network learning network. In order to use the pre-learned data as the next prediction data, it is a model that can infer the next motion by using the image itself between motions as an input variable.

2

2 is an algorithm block diagram of a motion prediction processor.

3

3 is a general hardware configuration diagram.

[Figure 4]

4 shows a motion vector.

5

5 is a shape recombination and motion tracking scene of an object.

Claims (5)

A TV remote control that extracts motion using predictive neural networks and uses it as control of a device. A device that divides and learns the expected motion pattern into the image itself when analyzing a specific image and uses it as data. Method of recombining each shape in the image and recognizing it as a motion pattern Device that can control the device by recombining each shape in the image and moving the recombined shape based on this How to use predictive algorithms when controlling your device with video.

Images