TW200521813A - Neural network correcting method for touch panel - Google Patents

Abstract

There is provided a neural network correcting method for touch panel. The touch panel is divided into plural blocks, and each block is configured with a correction point. The selected coordinate value obtained by measuring each selected correction point is used as an input signal, and the original coordinate value of each configured correction point is used as an output signal. Based on the relation between the input signal and the output signal, neural learning algorithm is employed to generate the block weighted parameter and pressure shift parameter for setting up update formula , furthermore use neural operation algorithm to determine the corrected coordinate value after calibration, thereby increasing the accuracy in selecting the touch signal.

Description

200521813 发明 Description of the invention: [Technical field to which the invention belongs] The present invention relates to a neural network correction method such as a touch panel, and more particularly to a correction formula for weight parameters and bias parameters obtained by using a neural-like learning rule. Simplify the correction method of neural network operation method to improve the accuracy of touch signal. [Previous technology] According to the press, there are many related products on the market using resistive touch panels. These touch panels themselves have a heterogeneous structure, and the panel touch panel has two different components. The relative coordinates of are not exactly the same. Furthermore, the characteristics of products made in the same process are also different, so it needs to be corrected to improve the accuracy of the touch signal. At this stage, the calibration method used in the inspection panel is mostly used to calibrate the calibration matrix through several minerals, and then the calibration operation of the correspondence relationship between the points, such as National Patent Bulletin Nos. 2 8 2 1 2 3 and 2 9 5 6 4 7 fresh. This operation requires precise trigonometric operations. When there are many selected points, the entire computing architecture will be complicated, and there will be a problem of slow calculation processing speed. If there are fewer selected points, there will be a large error, so it is necessary to improve it. [Summary of the Invention] The main purpose of the present invention is to provide a neural network correction method such as a touch panel that can improve the accuracy of the control signal and solve the problem of the sense of meaning. In order to achieve the purpose stated in the present invention, the present invention divides the touch panel into several blocks, and sighs at least one correction point in each block. The coordinates of the selected coordinates obtained by clicking each correction point are used as input. And ^ (the original coordinate value of each correction point (displayed on the display) is used as the output signal 200521813 The relationship between the wheel signal and the wheel minus is calculated by the 3 Gods method, and it is based on the input signal and the wheel output signal. _ 纟 __ The parameters and bias parameters of each health block, and the neural network-like material is used to describe the value of the red wire, and the accuracy of the signal is high. The other goals and advantages of this ignorance are different, It is not difficult to obtain an in-depth understanding from the detailed description and drawings of the following embodiments. ⑽, the present invention allows some differences in the arrangement of a certain document, or the arrangement of other documents, but the selected embodiment is used. Then in this manual, it will be explained in detail, and its structure will be shown in _. [Performance] The touch panel of the present invention is divided into several blocks by the network correction method, and is used in Every-District Finance Sets at least-the righting point The coordinate value is used as the rotation position, and the corrections set by Yuanyuan Laipu are used as the output signal, and Renshenhong's online learning method calculates the relationship between the input signal and the output signal, and trains the student according to the relationship between the input signal and the rotation signal. The weight parameters and bias parameters of each block are then implemented using a neural network-like algorithm to formally calculate the coordinates of the school money. Examples are described below: The first embodiment is a lion hall ab software simulation. _ The control panel i is divided into a matrix of five by five and a total of five blocks, and a correction point η is set for each block, as shown in the second figure. Please refer to FIG. 2, which uses the present invention and the like. Neural network (Neurai's correction method _, touch _ in operation, the first step is to correct the money line, if it is normalized, that is, enter the 25-point correction to perform the God-like method Lai Ji parameters (take away Partial to 200521813 pressure parameter (b), get its modified formula to update the weight parameter (Wx, Wy) and bias parameter (b) of the neuro-anamorphic method, and find X η = Wx * x + b Υ η —Wy * y + b Under ^ operation mode, no correction is performed, and the input signal of the selected coordinate value is directly substituted into the positive The formula uses _ to calculate the corrected coordinate value through network algorithm calculation. Enter the twenty-five correction points τn with a positive k H, and measure the coordinate Pη of each point as the input signal. The original coordinate value of each correction point τn is used as a riding signal. In this embodiment, the original coordinate value of each correction point τn is set to display the coordinate value of the correction point of riding _; please refer to FIG. 3 for visual indication. The coordinates of each point, the original coordinates of each correction point, and the coordinates after each correction. The coordinates of each point P η (without a random error amount of it 0.05) are as follows: P1 = [0,0]; P2 = [1Α1]; P3 = [2Α2]; PΦ = [3? 0.1]; P5 = [4,0] P6 = [0 · 1,1]; P7 = [l, l]; P8 = [2, l]; P ^ [3 , L]; ΡΚΚ3.94] P11 = [0.2? 2]; P12 = [1? 2]; P13 = [292]; P1Φ = [3? 2]; P15 = [3.8,2] P16 = [01? 3]; P17 = [1? 3]; P18 = [2? 3]; P19 = [3? 3]; P20 = [3.9? 3] P21 = [0,4]; P22 = [1,3 · 9 ]; P23 = [2,3 · 8]; P24- [3,3 · 9]; P25 = [4,4] The original coordinates Tn of each correction point are as follows: Τ1 = [〇, 〇]; Τ2 = [1 , 0]; Τ3 = [2,0]; ΤΦ = [3,0]; Τ5 = [4,0] Τ6 = [0,1]; Τ7 = [1,1]; Τ8 = [2,1] ; Τ9 = [3,1]; Τ10 = [4,1] ΤΙ 1 = [ 〇, 2]; Τ12 = [1,2]; Τ13 = [2,2]; Τ1Φ = [3 ^]; Τ15 = [4,2] Τ16 = [〇3]; Τ17 = [1,3]; T18 = [2,3]; T19 = [3? 3]; T20 = [4,3]; T21 = [0,4]; T22 = [1,4]; T23 = [2,4]; T2Φ = [ 3,4]; Τ25 = [4,4] Then calculate the relationship between the input signal and the output signal with a neural network-like algorithm. The MatLab program is as follows: for i = l: l: 25 200521813 // 25-point measurement coordinates, With random error amount of + »0.05 end W = [00]; tF = [0]; PWFb] = l_—a (P, T (:, l) ,, W, b, l, OO ^ // use] VMab's existing leak_a〇ftmctiai, find the weight of X and the bias value (Fb)

Xn = FW * P, + Fb; // Calculate X coordinate after correction W = [00]; b = [〇]; [FW ^ = 1 identify a (F, T (:, 2) 丨, W, b, Satoshi ^ // Use Matlab's existing leak_a〇fonction to find the weight of Y (FW) // and the bias value ㈣

Yn = FW * P, + Fb; // The γ coordinate El = sim after calculating the positive cross (both iy-T (:, l)). A2 > + sim (both 2) ¾ 〃Calculate the value of the selected coordinate P The amount of error from the correction point τ E ^^ sumCCXn-my ^ + sumCCYn-TOV ^) The error amount of the standard value (5 ^ Υη) and the correction point τ is similar to that of the neural network (Neural Network) based on the input signal and the rotation signal The relationship training learns the weight parameters (Wx, Wy) = [0.0009 L0373] and the bias parameter ⑼ = -〇〇598 for each block, and the correction formula is:

Xn = Wx (0.0009) * X + b (-0.0598) ''

Yn ^ Wy (1.0373) * Y + b (-0.0598) Multiply the selection coordinate value of fifteen points by each weight by the weight parameter (Wx, Wy) = [〇〇〇〇9 L0373] 'Add bias The number of dads (b) = _〇〇598, to obtain the new coordinate (χη, γη) of the new twenty-five point correction data is as follows: The running coffee household [〇side, side 39] wears ώώΜ〇 and 〇〇 05 clams; 200521813 (Xη3, Υη3: K1 · 9818, 0.1149]; (Xη4, Υη4) = [3 · 0615, 0 · 0585]; Qing, Υη5Μ4 · 574, Le 824] _6, Υη㈣. 0535, 0.98 卬 (χη7, Υη7 ) = [〇 · 9439, 1.05155]; (Χη8, Υη8) = [2 · 〇492,0 · 9421]; (χη9, Υη9) = [3 · 0181, 0.995]; (χη10, Υη10) = [4 · 0110,1 · 0158]; (Xnll, Ynll > = [0.1359,2.0011]; CXnl2, Ynl2) = [0.995U0346]; (^ 13, Υη13Η1.9706α.9772]; (× η14 ^ _5, Υη15Μ3 · 8517 ^ _; (Xη16, Υη1 ^ · _, 3.0837]; _7, Υη17M0 · 9357, 3.015]; _8, Υη18 ^^^^ _9, Υη19 households [3 · 0172, 3.0324] (× η20, Υη20 households [3 Zhe Qing1, Υη21Η 德 90,4 · _]; _ 2, ώιώ2Η〇 · _, 4.0111]; _3, Υ ^ Ηΐ · 3_π (χη24, Υη24Η3 · 0506,4 · _ ; (Χη25, Υη25 > = [4 · 0459,4 · _, ·, will be One-point selection of the coordinate Pη (including a random error amount of ± 0.05) is added to the square of the original coordinate of the correction point, which is subtracted from η. The resulting error amount is 0.2934. The corrected coordinates (χη, Υη) and correction for each point The squares of the original coordinates of the points, Tη, are subtracted and added. The resulting error amount is 01951. It can be seen that the corrected coordinates (χη, Υη) improve the accuracy by 33.5%. The touch panel immediately after the father is in the normal operation mode. There is no need to perform further corrections. When the user is counting money, the chain pair that is gradually obtained will be read instead ^, _ face, and _ road show algorithm. Different correction formulas to obtain the pure chain (χη, Υη) to improve Refer to Figure 4 and Figure 5. Refer to Figure 4 and Figure 5. The board is divided into A, B, and X-axis by five coordinate points according to the coordinates of the Υ-axis. Second embodiment of the invention

To the new Xn, Yn values. It consists of five blocks C, D, and E on the touch surface, and each block has 25 correction points according to the coordinates. The parameters (b) are divided into five blocks to update the neural-like evolution. In normal operation mode, no calibration is performed. The measured block is replaced by the weight parameter of the block. 'Using a neural-like algorithm to update the implementation of the simulation of a touch panel. Matrix operation Matlab 15 Correct the points, and perform the calibration. 200521813 Timing, first select the twenty-five correction points Tn one by one, and measure the selected coordinates h as input signals, and use the set original coordinate values of each correction center as the riding signal. Satoshi please refer to Figure 6. In the figure, g π has the loyalty coordinates of each point, the original coordinates of each correction point, and the corrected coordinates. The coordinates of each point are as follows: The five points in the A block are: A1 = [0 0]; A2 = [l 0.1]; A3 = [2 0.2]; A4 = [3 0.1]; A5 = [ 4 0]; The five selected coordinate points in block B are:

Bl = [0.1 1]; B2 = [1 1]; B3 = [2 1]; B4 = [3 1]; B5 = [3.9 1]; The five selected coordinate points in block C are:

Cl = [0.2 2]; C2 = [1 2]; C3 = [2 2]; C4 = [3 2]; C5 = [3.8 2]; The five selected coordinate points in the D block are:

Dl = [0.1 3]; D2 = [l 3]; D3 = [2 3]; D4 = [3 3]; D5 = [3.9 3]; The five selected coordinate points in the E block are: E1 = [0 4]; E2 = [1 3.9]; E3 = [2 3.8]; E4 = [3 3.9]; E5 = [4 4] The original coordinates of each correction point are as follows: 5 corrections in the Α block The point coordinates are: ΤΑ1- [0 0], ΤΑ2- [1 0]; ΤΑ3 = [2 0]; ΤΑ4 = [3 0]; ΤΑ5 = [4 0]; The coordinates of the 5 correction points in the Β block are: : TB1- [0 1]; TB2- [1 1]; TB3 == [2 1]; TB4 = [3 1]; TB5 = [4 1]; The coordinates of the five correction points in the C block are: TC1 1 [0 2]; TC2— [1 2]; TC3 = [2 2]; TC4 = [3 2]; TC5 = [4 2]; The coordinates of the 5 correction points in the D block are TD1 = [ 0 3]; TD2 = [1 3]; TD3 = [2 3]; TD4 = [3 3]; TD5 = [4 3]; The coordinates of the 5 correction points in the E block are: ΤΕ1 = [0 4] ΤΕ2 = [1 4]; TES3 = [2 4]; TES4 = [3 4]; Τ5 = [4 4] Then use a neural network-like algorithm to calculate the relationship between the input signal and the rotation signal, 1 200521813 Α (ι ,: Household ACvMW—ayZO; // 5-point measurement coordinates in area A have a random error amount of + Ό.05 aid W = [00]; M〇]; [AmAbX] ^ eam_a (A; TA (: a ); W? BaA〇l); // Use Matlab's existing leak_a〇fonction to find the weight of area A // (AWX) and bias value (AbX)

XiiA = AWX * A ’+ AbX; // Calculate the X coordinate of the Λ region after correction W = [00];

[AWYAbYHeam—aOTA ^ WW, 0.011); // Use] \ 131: 1313 Existing 1 £ 3111_ ^ 〇 & 01〇00〇11, find the weight of the eight regions // (AWY) and the bias value ( AbY)

YnA = AWY * A, + AbY; // Calculate the Y coordinate of the A area after the calibration i fori = l: l: 5 B (i, :) FBCi;) + (4) ^ * iand (l >20; // 5-point measurement coordinates in area B, with random error amount + W.05 end W = [00];

Μ〇]; [BWXBbXHeanuiOVIBCaXWWAOl); // Use Matlab's existing leakjaOfbnction to find the weight of region B // (BWX) and bias value (BbX)

XnB = BWX * B '+ BbX; // The white ^ jX coordinate of area B = [00]; M〇l; PWBbYHeam_a (&; IB (: 2) ,, W, b, l, 0.011); // Use Matlab's existing leak_a〇fonction to find the weight of area B // (BWY) and bias value (BbY) 11 200521813

YnB = BWY 氺 B ， + BbY; // Calculate the Y coordinate of the B area after correction fori = l: l: 5 C (i, :) FCCi ?: M4 ^ Tand (iy20; // 5-point measurement in the C area Coordinates with a random error amount of + -0.05 end W = [00]; b = [0]; [CMX (^ XHeam_a (C, TC (:, l) ,, W, b, l, 0.01); // Use Matlab's existing leak__a〇fonction to find the weight of the C region X // (CWX) and the bias value (CbX)

XnC = CWX * C + € bX; // Calculate the X coordinate of the C area after correction · W = [00]; M〇]; [CWYCbY] = leam—KCVTQ ^ WWAOll); // Use the existing Leam_a〇fonction of IvMab, Find the weight of C region Y // (CWY) and bias value (CbY)

YnC = CWY * C + € bY; // Calculate the Y coordinate of the C area after correction fori = l: l: 5 DCvWv juice (-收 — 卿;

// 5-point measurement coordinates in the D area, with randomness of + _〇 · 05: error amount and W = [00]; M〇]; [DWXDbXHeam_a (D ,, TD (:, l) ,, W, b, l, 0.01); // Use the existing lambLaofbnction of Matlab to find the weight of D area X // (DWX) and bias value (DbX)

XnD = DWX * D ， + DbX; // Calculate the X coordinate of D area after correction w = [oo]; b = [0]; PWY DbY] = leam_a (D; IIX: 2); W, b? L A 〇l 1); 12 200521813 // Use Matlab's existing leak-afbnction to find the weight of D area X // (DWY) and bias value (DbY)

YnD = DWY * D ， + DbY; // Calculate the Y coordinate of D area after correction fori = l: l: 5 Ε (ι;) = Ε (ι,:) ^ (-1) ^% ιηά (1 &gt;20; // 5-point measurement coordinates in the Ε region, with a random error amount of + <0.05 aid W = [00];

pVv ^ aXHeam-a (p ,, TT ^ :, l) ,, W, b, l, 0.01); // use] ^ 1 ^ existing 103111- &amp; 〇: [1 | 1] 〇11〇11 To find the weight of £ 1 (/ (EWX) and bias 郷 bX)

XnE = EWX * E + EbX; // Calculate the X coordinate of the E area after correction W = [〇〇]; LEWYEbYHeam—aOVIEOWbaAOll); // Use] \ ^ 1〇13 Existing 10311 \ _ ^ 〇 &amp; 111〇 11〇11, find £; the weight of the area y / (pWY) and the bias value (EbY)

YiMWE ’+ EbY; // Calculate Y coordinate of E area after correction

ElA = sum ((A (: a),-TA (: a)) A2) fsum ^ // Calculate the amount of error between the selected coordinate values A1 ~ 5 in area A and the correction point TA, EZA ^ sunXP &amp; A-TACUD / ^ sumCOfnA-TAdZ) // Calculate the amount of error between the miscellaneous positive touch target value (XnA ^ YnA) and the correction point TA ΕΙ ^^ ((β (:. ^-ΎΒ (:. 7ϊ) Υ ^^ // Calculate Beta The amount of error between the coordinate values B1 ~ 5 in the area and the correction point ΠB EZB ^ sun ^ CXnB-TB ^ yv ^ fsum ^ ynB-TB ^ y ^) // Calculate the value of the tooth ticket after the positive cross in Area B (XηΒ, Υηβ The amount of error from the correction point tb

ElC ^ sam ^ CCay-ICCa ^^ sum ^ 13 200521813 // Calculate the amount of error between the C-point selection coordinates Cl ~ 5 and the correction point TC in the C area E2G = sum ((XnC -TQ ^ V ^ sumiXYnC-TQ ^ V ^) // Calculate the amount of error between the corrected coordinate value (XnQYnC) of the C area and the correction point TC E1D = with (both 1) items :, 1)), 2) + surface (financial office takes 2)) production 2) // Calculate the amount of error between the coordinate values D1 ~ 5 in the D area and the correction point ID EZE ^ sumPhlXrDCAV ^ sum ^ ynDTD ^)) / ^) The amount of error between the coordinate values and the correction point ^^) E1E = coffee (both // calculation The error amount of the coordinate value E1 ~ 5 and the correction point TE in the E area. Leg == ((5 &amp; ^ 1 £ (:, 1)) · 8 2) + · 8ιπιι ((% Ε-ΤΕ (:; 2) ) · Α2) // Calculate the amount of error between the coordinate value ⑶ and the top of the correction point after the correction in the Ε region Ε1 = Ε1Α · ΐ1Β + Ε10Φ1ΓΗΕ1Ε The amount of error between the standard value and the correction point ΙΕ

E2 = E2A + E2B + E2C + E2EHE2E The error amount of the 彳 M standard value and the correction point TE The neural network (Neural Network) trains and learns according to the relationship between the input signal and the rotation signal to obtain the weight parameter (Wx, Wy) and bias parameter (b): weight parameter of the X coordinate of block A (Wx) = [0.9824 0.0373] bias parameter (b) = 0.0160; weight parameter of the Y coordinate (Wy) = [〇〇] bias Parameter (b) = 〇; Weight parameter (Wx) of block X coordinate in block B = [1.0197-0.0385] Bias parameter (b) = -0.0317; 200521813 Weight parameter of Y coordinate (Wy) = [-0.0049 0.4940] bias Pressure parameter (b) = 0.5061; weight parameter of X coordinate of block C (Wx) = [1.0489 -0.0431] bias parameter (b) = -0.0358; weight parameter of Y coordinate (Wy) = [-0.0078 0.8031] bias Pressure parameter (b) = 0.4093; weight parameter of the X coordinate of block D (Wx) = [1.0159 -0.0183] bias parameter (b) = -0.0047; weight parameter of the Y coordinate (Wy) = [-0.0114 0.8995] bias Pressure parameter (b) = 0.3099; weight parameter of the X coordinate of block E (Wx) = [0.9833 0.0001] bias parameter ⑻ = 0.0181; weight parameter of the Y coordinate (Wy) = [0 · 0159 0.9442] bias parameter ( b) = 0.2845; the coordinates of the measuring points at 25 points A1 ~ 5, B1 ~ 5, Cl Each of ~ 5, Dl ~ 5, and El ~ 5 is based on its relative block multiplied by the block's weight parameter (Wx, Wy), and the bias parameter (b) is added to obtain a new twenty-five point corrected The coordinates are as follows: Block A (XηΑ1, ΥηΑ1) = [-0.0169 0]; (ΧηΑ2, ΥηΑ2) = [1.0514 0]; 200521813 (ΧηΑ3, ΥηΑ3) = [1.9543 0]; (χηΑ4, ΥηΑ4) = [3.0112 0]; (ΧηΑ5, ΥηΑ5) = [3.9449 0]; Β block (ΧηΒ1, ΥηΒ1) = [0.0251 0.9963]; (ΧηΒ2, ΥηΒ2) = [0.9897 1.0153]; (ΧηΒ3, ΥηΒ3) = [1.9481 0.9799]; (ΧηΒ4, ΥηΒ4) = [3.0326 1.0073]; (ΧηΒ5, ΥηΒ5) = [3.8706 0.9632]; Block C (XnCl, YnCl) = [0.0532 1.9867]; (XnC2, YnC2) = [0.9442 2.0216]; (XnC3, YnC3) = [1.9673 1.9934]; (XnC4, YnC4) = [3.0323 L9985]; (XnC5, YnC5) = [3.8540 1.9784]; Block D

(XnDl, YnDl) = [0.0209 2.9886]; (XnD2? YnD2) = [0.9989 3.0352]; (XnD3, YnD3) = [1.9477 2.9640]; (XnD4, YnD4) = [3.0287 3.0106]; (XnD5, YnD5) = [3.8793 2.9437]; E block (XnEl, YnEl) = [-0.0040 4.0392]; (XnE2, YnE2) = [1.0239 4.0043]; (ΧηΕ3, ΥηΕ3) = [1 · 9648 3.8843]; (XnE4, YnE4) = [3.0127 4.0577]; (XnE5, YnE5) = [3.9513 4.1244] Add the squares of the subtraction of A1 ~ 5 and TA1 ~ 5 at each point of block A, and the error amount of the obtained error is 0.0778 '(XnAbu 5, YnA [5] Add the sum of squares subtracted from TA1 ~ 5, and the error amount of the obtained error is 0.0082. Add the sum of the squares of subtracted B1 ~ 5 and TB1 ~ 5 at each point of block B, and the amount of mistakes of the resulting error is 0.0370, (XηM ~ 5, YnBl ~ 5) and the square of the subtraction of TB1 ~ 5, the error amount of the obtained error is 0.0233. The sum of the squares of the subtraction of C1 ~ 5 and TC1 ~ 5 of the mother point of block C is added. The Wu Shengdan error is 0.0735 '(XnCl 5, YnCl ~ 5) and the square of the subtraction of TC1 ~ 5 is added. The voice of Wu Chadan is 0.0305. Add D5 at each point of block D and subtract the square of TD1 ~ 5 to get the error. 'Den, 0.0305, (XnDl ~ 5, YnDl ~ 5) and add the square of TD1 ~ 5. So, the loss of 16 200521813 of 5 is 0.0245. Add the square of the subtraction of E 5 and TE 5 at each point of the E block, and the error amount of the resulting error is 〇ΌόΠ (XnEl 5, YnEl 5) and the square of the TE 5 subtraction. The amount of error is 0.0381. 〇 Add the wrong buttons of the five blocks, uncorrected Yilin difference _ = () • belong to, after correction # total error amount Ε2 = () · 1246, it can be seen that the corrected point increased by 56 · 5 % Accuracy.

After correction, the difficult panel will not be corrected in the normal operation mode. When the user clicks the touch panel, the age of the Japanese-style system is determined by the weight parameters of the block (Wx, Wy). ) Touch her _ Jianli Zhixiu, __ return algorithm to calculate the corrected chain (Xn, Yn), bet high touch signal accuracy. In addition to the above-mentioned correction formulas of the real turning butterfly and the transfer operation, the red silk of the material layer type can also be modified by a neural-like algorithm operation. For example, the correction formula is established as ...

Yn = Wyl * Y2 + Wy2 + b \ Find ^ is called: the layered formula (21㈣ 之 _calculation E, detailed miscellaneous classics learning method to look at the machine, b parameters and missing parameters are edited to form a multilayered modified formula. Τ, said, this According to the invention, by setting the touch panel to be divided into several blocks, and dividing the points in each block, the measurement points are used to select corrections, and then the coordinate values of the Shanghai ^ 〃 X wait point are used as input signals. The original coordinate value of the 疋 σ correction point is used as the round-out percent, and the number of the pass # w, and the Yayi round-in signal and round-out signal% are calculated using the neuron-like f rule, and the corrected value is obtained. Change &gt; number and test parameters to establish a modified I value of the quasi-cancer rate of the touch signal selected from the south point. 17 200521813 j The disclosure of the above-mentioned examples is used to explain the present invention, not to limit the present invention. Therefore, any change of numerical values or replacement of equivalent components should still belong to the scope of the present invention. From the above detailed description, the skilled person will be able to achieve the aforementioned purpose, and it has already complied with the provisions of the Patent Law. , I filed a patent application. [Brief description of the drawings] Le 1 is the first embodiment of the present invention The touch panel is divided into twenty-five blocks and provided with twenty-five correction points. The second diagram is a flowchart of a neural network correction method such as the present invention. The third diagram is the first embodiment of the present invention. Schematic diagram of the coordinates of the selected coordinates, the original coordinates of each correction point, and the coordinates after correction. Figure 4 ^ This task will be described in the first embodiment. The touch panel is divided into five blocks with twenty-five correction points. Figure 5 is a neural network correction method such as the second embodiment of the present invention. Figure 6 is a map (Ding, the coordinates of each point selected in the second embodiment of the present invention, the original coordinates of each correction point, and the coordinates after correction. Location diagram 18

Claims (1)

200521813 Scope of patent application: 1. A neural network correction method such as a touch panel, which divides the touch panel into several blocks. Sub-per-block set at least-healthy points, measurement points Select the point coordinates obtained by each calibration point as the input signal, and set the original calibration coordinate values as the output signal. Use the neural network-like learning method to calculate the relationship between the input signal and the rotation signal. The relationship between the input signal and the output signal is trained to obtain the weight parameter and the bias parameter to establish a correction formula, and a neural-like algorithm is used to calculate the correction formula to obtain the corrected coordinate value. 2 · A neural network correction method such as a touch panel as described in item # 1 of the patent scope, wherein the touch panel is divided into twenty-five blocks by a five by five matrix, and each block is -Correction points are set separately in order to obtain correction formulas for correction by performing neural-like learning rules respectively. 3 According to the neural network correction method such as touch panel described in Item 1 of the declared patent scope, where the smart panel is divided into five blocks according to the gamma axis of its seat ^, each block is based on the X car of the coordinates She was on time by Xiang Dian 4 5 and corrected by the correction formula of five blocks obtained by splitting the job and supplementing the neural learning method. 4. Apply for a neural network correction method such as the touch panel described in # 1 patent fesl 帛 1, where the network learning method calculates the relationship between the input signal and the rotation signal, and trains according to the relationship between Γ M w and output money Learn to get weight parameters and bias parameters to establish