TWI352805B - - Google Patents

Description

1352805 IX. Description of the invention: [Technical field of the invention] The present invention is a method for measuring the brightness of a test object, and a kind of neural network as a core can be used together with a digital camera to establish a surface luminance distribution of a test object at a constant speed. A two-dimensional glow measurement method for images. [Prior Art] According to liquid crystal display (Liquid Crysta丨Disp|ay, LCD), it has the advantages of thinness, light weight, low power consumption, no radiation pollution, and compatibility with semiconductor process technology. One of the industries, in which the backlight module (BLM) is one of the key components for providing LCD backlights. Since the liquid crystal display must rely on the backlight module to emit light, the quality of the backlight module directly affects the performance of the liquid crystal display, and luminance is one of the indicators for evaluating the quality of the backlight module. Radiant Imaging has launched the first CCD-based luminosity measurement device 20 years ago. In addition, the Japanese company | System and Mi no丨ta also have a full-scale luminance color measurement device; In China, companies such as Arma Optoelectronics are developing related products. The working principle of the above-mentioned measuring equipment is that the image of the surface to be tested is first captured by the CCD, and then analyzed by the software developed by each company to obtain the required two-dimensional luminance output.

Radiant Imaging pointed out that its color measurement device can quickly measure the color coordinates and luminance of the display, mainly using its developed CCD photo/color measurement system device' with its developed software 4 1352805

Prometric ϋ# ^ χ , 均十八曰ι τ贝枓 for analysis. However, the measurement equipment developed by the aforementioned company, the mussels, is based on the fact that some manufacturers are more economical and more cost-effective, and the silk selection is subject to questioning. However, the reliability of the results is still [invention] The main objective is to provide a low-cost and fast-processing method that can be used to measure the opposite side of the production line (4), thereby increasing the measurement speed on the production line and reducing its cost. The technical means for achieving the foregoing objectives are such that the method comprises the steps of: a. applying different operating voltages to one side of the light source and taking the luminance values at different operating voltages at the same designated position of the surface light source; b. Applying different working voltages to the surface light source, and using cCD to capture images at different working voltages at the same specified position of the surface light source, and respectively converting them into a gray scale image; c. The original gray scale value is taken out from the specified position of the image, and the original gray scale value has a proportional relationship with the luminance value; d. according to the original gray scale value and the luminance value corresponding to the complex array, the curve fitting method is used to guide the correlation function relationship e. Define a plurality of feature point positions in the surface light source, and measure each feature point position to take the correct luminance value, and then reverse the function relationship formed by the curve fitting method in the above step d a corrected gray scale value of the complex feature point position; 5 1352805 f. The image of the surface light source is respectively captured by the CCD, and then respectively taken at the position of the plurality of feature points The uncorrected gray scale value; g. obtains the complex array "eisk value gray scale correction coefficient" according to the uncorrected gray scale value and the corrected gray scale value corresponding to the foregoing complex array (gray scale correction coefficient = corrected gray scale value / Uncorrected grayscale value); h. Use the neural network to form a grayscale correction matrix according to the aforementioned "eigenvalue grayscale correction coefficient"; 丨 correct the image obtained by the CCD according to the grayscale correction matrix described above. The "Grayscale Correction Matrix" and the fitted curve of the corrected grayscale value and the luminance obtain a simulated two-dimensional luminance distribution. The above method can be used together with the advantage of the digital camera to quickly take images, and quickly obtain a two-dimensional plane or even The luminance distribution of the curved surface not only ensures the reliability of the metric measurement, but also greatly reduces the cost β. [Embodiment] The present invention uses a digital camera to obtain an image of the luminance distribution of the surface to be measured, and the charge coupling component (Charge) Coupled Device (hereinafter referred to as CCD) is the image capture component used by most digital cameras. In traditional single-point measurement of luminance meter, the measurement point and its detector must be Straight orthogonal, in order to get the correct luminance value. When using a digital camera to take pictures, the working principle is as shown in the first figure, a uniform brightness light source is located in the object plane of the imaging lens' while the CCD is seated. Falling on the image plane of the imaging lens, the first ^ φ * - can not produce a uniform luminance source, and its luminance L can be calculated by 6 1352805 degrees E; where: Luminance (L) is defined in the unit light-emitting area Projection A, the luminous flux measured at the solid angle. Increase = $ ' where 牦 (1)

Θ is the angle η( cd/m2) ° between the direction of observation and the direction of the light-emitting surface. The unit of luminance is obtained by substituting the photometric formula. E E1=(10) * 1 dA, , άΦ I〇=dQ0 (2) (3) Therefore, the luminance L L0- d2° _dI 〇 and 0 cosordA0dn0 cosadA0 dA0 (4) are obtained by substituting equation (3) into equation (1). Since a point radiation source has the same radiation intensity in all directions, it emits in a limited solid angle. The radiant flux is Φ = ΙοχΩο (5) id (5) generation (2) illuminance and luminance relationship E _da〇x〇0) _Q〇d(I0) dA '蛑_ dAT = L. ^^ (6) Substituting the photometric formula to obtain E π d(D 3 E3=i dQ, (7) (8) Therefore, by substituting equation (8) into equation (1), the luminance L2 = ά2φ cosadAjdQ, cosadA2 (9) Since a point source has the same radiation intensity in all directions, its radiant flux emitted in a finite solid angle is 7 1352805 (10) Φ = I, χ Ω, so the comprehensive formula (9), 10) Substituting (7) can obtain the relationship between illuminance and luminance _dq, xQ,) Ω, άα.) 3 dA3 dA3 = L2 dA2 〇 -^-^cosa dA3 (11) Because the second figure can be seen ^^cosa, according to The solid angle definition can be obtained as ΩοΑΆ#2, so Ω, = Q0cos2a (12) can be obtained by substituting equation (12) into equation (11).

dA

E3=L2-^-Q0cos3cz (13) Because the normal of the 3 faces and the rays emitted by the private face are inferior to the angle of the a, the law of the cosine cosine is known (14) ^3=^2 T~^^〇 COs4a dA3 is a uniform luminance source because it is shown on the first figure, so l0=l2 so the final e3 e3 (15)

The equation (1 5) indicates that in addition to the illuminance on the horizontal axis, the illuminance values of other positions have a cos4a error, so when using the digital camera for planar luminance detection, the output image must be corrected. As for how to cooperate with a digital camera to perform a two-dimensional luminance knife with correction function, the specific technical means are as follows: As shown in the second figure, it is mainly provided with a test platform at one end of a reticle (1 〇). (1)), opposite to the other end is provided with a fixing seat (12) for erecting - digital camera or - luminance meter, 2 fixing seat ((7) and test platform (11) parallel to each other to ensure digital camera or The light source to be tested on the luminance meter test platform (1) is in an orthogonal relationship. The data of the digital camera and the luminance meter are sent to the software Mat丨I, and the difference can be handled by the mathematical method... 仃 numerical value, image, neuron, user interface setting. As for the vertical a > narrow " sigh 疋 ❶ 〃 〃 〃 〃 〃 〃 〃 仃 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 详 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃 仃In the embodiment, as shown in the fourth figure, nine feature point positions are defined on the surface light source, and the 荽 荽 兀 々 々 々 出

The brightness of the 5th feature point in the center of the light source, and the gentleman #&, by supplying different working voltages to the backlight module, respectively obtain different luminance values at the 5th feature point position. In the second step, the illuminating knife cloth for obtaining the surface light source is captured by means of CCD image capturing, wherein the CCD image capturing is based on the position of the fifth feature point as the output image center. The image obtained by CCD is sent to the computer to learn the #RGB three-dimensional matrix in the image through Matlab mathematics, and then use Mat|ab

The image processing built-in function converts the RGB three-dimensional matrix into a two-dimensional gray-scale value matrix to form an original illuminance distribution map, and then knows the original gray-scale value of the fifth feature point position. In the correction principle of the digital camera photographing value, the final derivation of the formula indicates that in addition to the illuminance on the horizontal axis (i.e., the fifth feature point position), the illuminance values at other positions have at least one (7)^^ error. Therefore, it is assumed that the original grayscale value of the fifth feature point position does not need to be corrected, and the curve can be matched with the actual luminance, so the luminance value of the fifth feature point position is theoretically proportional to the original grayscale value. relationship. Then, by supplying different working voltages to the backlight module, the corrected gray scale values of the complex array are respectively obtained at the fifth feature point position. After 9 1352805 • * repeats the first step and the second step, the data of the corrected gray scale value pair (four) degree value can be obtained, and (4) the curve fitting method leads the related function relationship.帛Three steps' use the luminance meter to measure the luminance of the surface light source. In this step, we measure the luminance of more than 9 feature points on the surface light source; use the second step curve to form the functional relationship, reverse The corrected grayscale value of the feature point position above 9 points.

In the fourth step, the working voltage of the previous step is fixed first, and then the illuminance distribution of the light source is taken by the CCD image capturing method; wherein the CCD image is taken as the output point of the 5th point of the image value to obtain 9 points. The uncorrected grayscale value of the above feature point position. According to the third step and the fourth step, 9 sets of "eisk value gray scale correction coefficient" (gray scale correction coefficient = corrected gray scale value / uncorrected gray scale value) are obtained, and then the nerve-like correction coefficient is used to use the nerve-like The network forms a grayscale correction matrix and uses this matrix to perform the calibration procedure required for the digital camera to take values. Therefore, using the above-mentioned "gray scale correction matrix" and "corrected Qin & fitting curve and luminance fitting curve" can be obtained - simulated: dimensional luminance distribution 〇 following the "step fitting" proposed for the above steps The method puts forward the explanation: The so-called "curve_f_g" means that a set of discrete data is represented by an approximate curve equation. The mathematical model usually established by curve fitting is "single input/single output". (Single input Sing output 'short for S丨S0)' so its characteristics can be represented by a curve. Assume that the data obtained by observation is as shown in the fifth figure. The table says that 10 1352805 clearly uses the voltage supplied to the backlight module from i〇0V~12 ov, and each time interval 〇2V' The characteristic point luminance value corresponds to the value of the central feature point of the backlight module obtained by the CCD photographing (gray scale value), both of which are vectors of length 10, assuming that the relative luminance value of the gray scale value is 〇~255, that is, A mathematical model must be established based on the above information and predicted based on this model. According to the data listed in the fifth figure, use Matlab's curve fitting ΤοοΙΒοχ to simulate the curve of the luminance value and the corrected grayscale value, and set the curve with the corrected grayscale value as the input 'luminance value to turn out'. Smoothing Spline is a suitable type of curve. Then further explain how the gray-scale correction coefficient forms a gray-scale correction matrix using a neural network: according to the literature, an ideal uniform surface light source obtained by photographing through a CCD (as shown in the sixth figure) is off-center. The closer the grayscale value is, the lower the grayscale value is, the farther it is from the center. If the above derivation formula is used, the error of cos4a can be inferred; but in fact, the uniformity of the surface light source itself is usually not so ideal, and the digital camera is in imaging. The error formed is not only the error of c〇s4a, so the present invention uses the inverse transfer type neural network to perform more complicated but precise correction work, mainly using the network to set several input vectors corresponding to the target vector, and set some parameters. Then, through the convergence test, the correct network relationship is obtained, which is applied into a gray-scale correction matrix, and the uncorrected gray-scale matrix of the original image (as shown in the seventh figure) can be converted into the corrected gray-scale matrix to form The correct two-dimensional luminance distribution image (as shown in Figure 8). The so-called "like neural network" is a computing system consisting of many artificial neurons (or processing units) connected by a height of 11 1352805. As shown in the ninth figure, a neural network usually has three layers. The artificial neurons of the hidden layer 'output layer β hidden layer may be one or more than two, each artificial neuron may be regarded as a separate processor, and the number of processors is two or more. Operate in parallel. The calculation process of the inverse transfer neural network is divided into a learning process and a recall process. The learning process consists of two phases: forward delivery and reverse delivery. Forward transfer starts from the input layer and passes to the hidden layer and then to the output layer. The layer is forwarded one by one and the output value of each processing unit is calculated by the nonlinear conversion function until the last layer of the network. The reverse transfer is transmitted backward by the output layer. This stage is to calculate the error and update the link value. The method corrects the difference between the output vector and the target vector, and sends the difference signals back to the network. Recalculate the link value and update the link value so that the output vector of the network is getting closer to the target vector, completing the training of the network. This learning process is usually performed in a training paradigm at a time until all training paradigms have been learned and network convergence is achieved. As for the recall process, the network uses the input vector input to estimate the output vector over the course of learning...:=. The following is a further explanation of the calculation method for modifying the link value in the reverse transfer network. ' In the inverted transfer network, the output value of the nth layer is the nth layer of the first neuron output value. The weighted cumulative value of the nonlinear function: yaj = f (net)) (21) where y; is the output value of the nth jth neuron, is the conversion function 12

丄 J JZ/OVJJ, (10)〗 is the n-1th layer i i also a god!权 The weighted cumulative value of the output value can be expressed as Λβί"=Σ^Γ'-6. | (22) % is the output value of the first neuron of the n_1th layer, and the % is the η·1 layer i nerves, (4) is the number of neurons in the first neuron. The chain value is reduced in the day; the J transmission path belongs to the supervised learning mode, and the training purpose is in the ^^ ” output to the gap between the two Therefore, the error function is usually used to represent the learning quality. The error function is: £=^Σκ~^)2 y' (23) : In the middle, the target output value of the jth neuron is the Jth of the output layer. The network of neurons gives the value, and the difference between the two is used to adjust the error function. Five dare to make it 'the best solution for the network to learn the steep slope method to search for the best solution' is the smallest error. ., square sum. Each time a training resource is input, the link value is slightly adjusted for the road. The magnitude of the J-direction adjustment and the error number are proportional to the sensitivity of the link value Δ^=-7ϋ (24) The product is the rate of the ', and its catch is corrected by the steepest slope method. The chain law of the micro-product knife can be used: (25) Key formula. This learning process is usually performed by inputting a training paradigm at a time, until the force is smashed into all the ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ ̄ In the aforementioned luminance distribution measurement step, according to the third step and the fourth step, the eigenvalue gray scale correction coefficient of 9 絚 or more is obtained, and in the inverse transmission type neural network application, the neural network training simulation part system "(Other luminance feature point X-axis value, maximum luminance feature point X-axis value)", "(Other luminance feature point Y-axis value, maximum luminance feature point Y-axis value)", "(Other luminance feature point original grayscale value) - Maximum luminance feature point original grayscale value)" as the three-in-one of the neural network, 'other luminance feature point gray-scale correction coefficient/maximum luminance feature point gray-scale correction coefficient" as a target vector, divided into training examples , test examples, verification examples. In the parameter setting: the hidden layer usually uses the layer. The more the number of processing units of the hidden layer, the slower the convergence, but the smaller the error value can be achieved, but too much does not necessarily reduce the error of the test paradigm, but increases the burden and time of execution, too little to react. The interaction between the input variables has a large error, and the more the number, the more complicated the network and the slower the convergence. The learning rate usually uses the maximum stable learning rate. Larger learning rates have larger network weighting corrections, which can approach the minimum value of the function faster. However, if the value is too large, the weighting value is overcorrected, causing the numerical oscillation to be difficult to achieve convergence. It is usually handled by an experience of ~1〇 or by automatically adjusting the learning rate. The above description can be used to understand the further technical content of the "curve fitting" and the "neural neural network" in the measuring method of the present invention, and the following steps can be further summarized in the foregoing measuring embodiment: a. Apply different working voltages and take the luminance values at different operating voltages at the same finger position of the surface light source; b. apply different operating voltages to the surface light sources, and use CcD to specify the same in the surface light source. Positioning the images at different working voltages and converting them into a grayscale image; c. respectively extracting an original grayscale value at a specified position of each grayscale image, the original grayscale value and the luminance value Having a proportional relationship; d. according to the original gray scale value and the luminance value corresponding to the complex array, using a curve fitting method to guide the correlation function relationship; e. defining a plurality of feature point positions in the surface light source, and applying a surface light source a fixed working voltage, and taking an image of the surface light source at the working voltage by the CCD, and then obtaining the operating voltage under the plurality of characteristic point positions respectively The uncorrected gray scale value of the surface light source image is further determined based on the function relationship formed by the curve fitting, and the corrected gray scale value of the complex feature point position is obtained inversely; f. according to the uncorrected gray scale corresponding to the complex array For the value and the luminance value, obtain the complex array "eisk value gray scale correction coefficient" (gray scale correction coefficient = corrected gray scale value / uncorrected gray scale value); g. use the neural network according to the aforementioned "eigenvalue gray scale correction coefficient" The network forms a grayscale correction matrix; h. corrects the image obtained by the CCD according to the grayscale correction matrix described above. 9 i. passes the grayscale corrected image through the “fitted grayscale value and luminance fitting curve” Converted into a two-dimensional luminance distribution image. The above method can be used to extract the image of the surface light source to be measured by the digital camera, so as to obtain a two-dimensional luminance analysis result with significant improvement reliability, since the analysis result may be interfered by noise, so it can be analyzed The result is filtering to eliminate the aforementioned noise interference. A feasible filtering technique is the median filter. The measurement of the aforementioned two-dimensional luminance distribution is defined as a "self-measurement, mode, which is mainly based on the gray-scale correction matrix of the gray-scale value and the luminance value measured by the light source itself to be measured, as a simulated two-dimensional The basis of the luminance analysis, therefore, when performing two-dimensional luminance analysis on different surface light sources, it is necessary to separately establish gray scale correction matrices for different surface light sources. _ In addition to the above modes, a fixed gray scale correction matrix can also be used. Simulated two-dimensional luminance analysis: The fitting curve and grayscale correction matrix of the corrected grayscale value and luminance are obtained for a backlight module by using the above-mentioned "self-measurement" distribution method. The grayscale correction matrix and the fitted curve of the corrected grayscale value and luminance perform a calibration measurement of the same batch or the same type of backlight module, which is defined as a "system amount" :, mode. [Simple description of the diagram] Fig. Schematic diagram of the error. 0 Schematic diagram of the location of the sign. First diagram: The principle of taking a digital camera is shown in the second figure. · The camera obtains the illuminance value. A FIG. System architecture diagram of the present invention the hard body of the fourth FIG. Laid the present invention is defined based on the surface light source

Dream-r* Figure is a table of related data for different luminance values and grayscale values generated by the operating voltage of the backlight module. Chu-X» Fig.: The image of the luminance distribution of the uncorrected gray scale using the CCD to extract the ideal uniform surface source. Λ|ϋτ | Fig.: Using CCD to capture an image of a backlight module (before correcting grayscale) Figure 8: Using CCD to capture the luminance distribution image of a backlight module (corrected grayscale via neural network) Rear)

Figure IX: Schematic diagram of a class of neural networks. [Main component symbol description] (1〇) remake frame (ή) test platform (12) fixed seat

17

Claims (1)

1352805 X. Patent application scope: 1. A two-dimensional glow measurement method based on a neural network, including: ^ ^ defines a plurality of feature point positions on one side of the light source, and applies different colors to the surface light source. The working voltage, and then the luminance values of the different operating voltages are taken at a certain feature point position, and the corrected grayscale values of the plurality of feature point positions are inversely obtained according to a function relationship formed by curve fitting; Applying an operating voltage to the surface light source, and capturing an image of the surface light source at the working voltage, and respectively obtaining an uncorrected grayscale value at the working voltage at the plurality of feature point positions; corresponding to the complex array The uncorrected grayscale value and the corrected grayscale value, obtain the complex array, and the eigenvalue grayscale correction coefficient", and then use a type of neural network to form a grayscale correction matrix; according to the grayscale correction matrix The image is corrected; a simulated two-dimensional luminance distribution is obtained by using the gray scale correction matrix and the fitted curve of the corrected gray scale value and luminance. The one-dimensional glow measurement method with the neural network as the core according to the first item of the benefit range, the uncorrected gray scale value and the luminance value are obtained by the following steps: _ applying different working voltages to the surface light source, and The surface light source takes the luminance value at different working voltages at the same decoupling position; applies different working voltages to the surface light source, and extracts images at different working voltages at the same buckle position of the surface light source And respectively converted to a grayscale image; 18 1352805 respectively extracts an original grayscale value at a specified position of each grayscale image, and the original grayscale value has a proportional relationship with the luminance value; according to the original gray corresponding to the foregoing complex array The order value and the luminance value are guided by the curve fitting method to guide the correlation function relationship. 3. A two-dimensional luminance distribution method based on the neural network as the core of the second application of the patent application scope The result can be further subjected to a filtering step. 4. As described in the third paragraph of the patent application, the neural network is the core. The dimensional measurement method is a median filter. The two-dimensional glow measurement method based on the neural network 3k U ... h~ described in any one of claims 1 to 4, wherein the luminance value is measured. The two-dimensional glow measurement method of the neural network #j, ''', the core neural network is a reverse transmission type neural network, as described in any one of items 1 to 4. The two-dimensional glow measurement method of the neural network fish, the core, and the core according to any one of items 1 to 4 is an image obtained by using a CCD to extract a surface light source, as described in claim 7 of the patent scope. Taking the neural network as the core and the two-dimensional 缂 measurement method 'The CCD is the CCDe in the exponential camera. XI, Fig.: Page 19