CN116994515B - Quick gamma correction method based on gradient descent - Google Patents

Abstract

The invention provides a method for rapidly approaching and finding out the corresponding voltage value conforming to the gamma curve condition by using gradient descent and assisting interpolation, fitting and other methods, and the method can be used for solving the gamma curve measurement, other similar problems can be solved by using the same method, and only the target and the initial input are needed to be provided.

Description

Quick gamma correction method based on gradient descent

Technical Field

The invention belongs to the field of display driving of display screens, and particularly relates to a rapid gamma correction method based on gradient descent.

Background

In display driving, the corresponding voltage value of the gamma curve is generally measured and given, which is a time-consuming task. Since three voltages are used to control the light emission intensities of the RGB sub-pixels in the display driving, the voltages and the light emission intensities do not correspond exactly to a certain known linear relationship, and the voltages are discrete values and finally react to the brightness when the gamma curve required by the display screen is the overall RGB to be lit, and the color coordinates x and y need to be kept as consistent as possible. Thus, each RGB three voltage value combination is involved to output the respective combined brightness (Lv), color coordinate x and color coordinate y, and the measurement is finally required to be a plurality of groups of RGB voltage value combinations, wherein the plurality of groups of combinations enable Lv to meet gamma curve characteristics as much as possible, and x and y are preset fixed values. The difficulty is that the three groups of voltages of RGB can independently influence Lv, x and y, and a plurality of objective functions directly have conflict and contradiction, so the problem can be regarded as a multi-objective planning problem, a great amount of time is consumed if an exhaustion method or manual measurement is used, the result deviation is larger if the linear regression fit is performed on the measured multiple groups of RGB data, and even if the two methods are combined, considerable time consumption still exists only when the exhaustion method is performed near the fitting result.

The invention provides an algorithm for gradient descent to measure, which can quickly approach the result to reach the measurement requirement, and can automatically obtain all the required RGB voltage combinations without manual intervention in the process after the required parameters are set.

Disclosure of Invention

The invention provides a rapid gamma correction method based on gradient descent, which comprises the following steps:

s1, inputting RGB voltage values, measuring brightness Lv and color coordinates x and y of a screen body, and comparing the brightness Lv, the color coordinates x and y with target values of theoretical targets Lv, x and y;

s2, comparing the results obtained in the step S1, and if the termination condition is met, selecting an optimal result; if the termination condition is not met, the RGB voltage value to be measured of the current target point is adjusted, and the step S1 is repeated until the termination condition is met;

s3, each time the RGB voltage value to be measured last time is adjusted, the RGB voltage value combination to be measured next time is updated by using a gradient descent algorithm, and the combination value is used as input to return to the step S1 to continue measurement;

s4, adjusting RGB voltage values to be measured of the next target point to obtain an optimal result; if other target points exist, respectively acquiring an optimal result of each target point, namely repeating the steps S1-S3;

s5, if each target point is measured, an optimal result is obtained, and the measurement is finished.

Further, the termination condition determining method in the step S2 is as follows:

when the measured values of the brightness Lv and the color coordinates x, y of the measurement screen body are close enough to the target values of the theoretical targets Lv, x, y, the measurement of the current round is terminated, and whether the brightness Lv meets the close enough requirement and simultaneously meets two limits is judged, wherein one limit is thatThe value of (2) is less than the first threshold, another limitation being a percentage limitation, i.eLess than a second threshold.

Further, judging whether the color coordinates x, y of the other two measured values are close enough to each other or not, wherein the two measured values need to be satisfiedA value less than a third threshold value, ">The value is less than the fourth threshold.

Further, adjusting the RGB voltage value to be measured at the current target point includes:

mean square error calculation, the formula isIn the formula +.>Is a target value->I is the sample number, n is the sample group number; taking three groups of measurement samples of brightness Lv and color coordinates x and y of the screen body, wherein n= 3,i =1 corresponds to Lv, i=2 corresponds to x, and i=3 corresponds to y;

the gradient is calculated by the formula:。

the gradient provides a direction and a step length, the calculated MSE value is stored in combination with the RGB voltage value for generating the group of data, so as to be used when the optimal result is selected later, and the gradient is used as a parameter to be transmitted to the gradient descent algorithm in the step S3.

Further, in step S3, the gradient calculated in step S2 is provided as a parameter to an optimizer, and the optimizer calculates a value updated this time;

the brightness Lv increases monotonically with the voltage change, the minimum value of the loss function is the minimum value, and the local optimal solution is the global optimal solution, so that convergence is accelerated according to positive and negative of the gradient:

s3.1, gradually increasing the learning rate when the gradient sign is not changed, and accelerating the convergence of the loss function;

s3.2, when the gradient sign changes, judging the relation between the current learning rate and the initial learning rate:

(1) If the current learning rate is higher than the initial learning rate, directly resetting to the initial learning rate;

(2) If the current learning rate is smaller than the initial learning rate, the current learning rate is further reduced when the current learning rate is smaller than the initial learning rate, the current learning rate is further reduced after the oscillation is started and the gradient direction is changed for a plurality of times.

Further, the optimizer is implemented by the following parameters: initial learning rateVariable a, direction d, learning rate->；

The first two parameters, namely the initial learning rate, need to be specified when the optimizer is initializedAnd the value of the variation parameter a, which is required to be specified in advance, the luminance gradient +_ among the gradients calculated in step S2 needs to be entered at each call>，Wherein->For the measurement of the brightness Lv +.>For the target value of the luminance Lv, the learning rate +.>Is updated according to the update of the update program.

Further, the learning rateThe updating steps of (a) are as follows:

wherein,for the current state learning rate, < >>Learning rate for the previous state; finding out learning rate->The post d will be updated to the sign of the brightness gradient and will learn the rate +.>Returning as an optimizer result;

obtaining the learning rateThen, the prediction update of the voltage combination to be measured next time can be started;

first, according to brightness adjustment, a rule is to use gradient of brightness Lv simultaneouslyTo uniformly adjust R, G, B:

wherein,for the last measured RGB voltage value of the current target point +.>，/>，/>；Is a voltage change value, 1 is taken if the voltage change value is between 0 and 1, and-1 is taken if the voltage change value is between-1~0; if the voltage change value has exceeded the input voltage value V, this is taken as +.>；

Secondly, the brightness is adjusted according to the chromaticity after the adjustment, and the rule is to use the gradient of the color coordinate xR is adjusted separately, using the gradient of the color coordinate y +.>The adjustment of G is carried out independently, and is specifically as follows:

the voltage is regulated for the second time according to the sequence, and the regulated voltage is obtained、/>、/>After the values are rounded off, the values are transmitted back to the step S1 as input to continue measurement until the termination condition is met, and the optimal result is selected.

Further, if the termination condition satisfied in step S1 includes, but is not limited to, the following conditions:

(1) When the gradient descends to reach the upper limit of a certain number of times;

(2) When the possible ranges have all been measured;

the mean square error MSE of all the measurement results in this round of measurement is counted and a group with the smallest MSE value is selected as the optimal result.

Further, in step S4, if there are other target points, the measurement is performed as follows:

s4.1, updating a target point;

s4.2, fitting the starting point of the next round;

s4.3, resetting the optimizer;

s4.4, starting measurement of a new target point again;

where S4.2 fitting the starting point of the next round requires the results of the last two rounds as a reference, and if there are no results of the last two rounds this step S4.2 is skipped.

Further, the fitting is specifically to calculate the inverse of the exponential term of the previous two rounds of results:

firstly, fitting calculation needs to be carried out on three RGB channels, binding point values of the previous two rounds are denoted as b1 and b2, voltage values are denoted as v1 and v2, currently calculated binding point values are denoted as b0, unknown number voltages needing fitting are denoted as v0, and a fitting formula of v0 is as follows:

second, for->Performing numerical detection, and if the value is not in the (0, 1) interval, directly using the voltage result of the last target point as a voltage starting value without adopting the fitting result as the voltage starting value of the next target point;

finally, only whenAbove 0 and below 1, the fitting result is considered to be authentic, using the fitting result v0 as the initial voltage value.

The method comprises the following steps: firstly modeling the problem, simplifying the problem from a multi-target planning problem to a single-target planning problem as much as possible, taking whether the mean square error of the targets Lv, x and y and the measured Lv, x and y after numerical normalization is closer to zero as a gradient descending target, simultaneously uniformly adjusting R, G, B by using the gradient of Lv, independently adjusting R by using the gradient of x, independently adjusting G by using the gradient of y, and the adjusting method is based on the objective rule of R, G, B on the influence of Lv, x and y.

After such a re-modeling, the original multi-objective programming problem of lv=f (R, G, B), x=g (R, G, B), y=h (R, G, B) becomes a single-objective programming problem in which the measured gradient causes the MSE to approach 0 as much as possible.

In the gradient descent algorithm, a learning rate and an optimizer are also needed to determine the step length of the current gradient direction, wherein the gradient descent is acted on a voltage and brightness curve, and the voltage and the brightness can be regarded as a monotonically increasing relationship.

Meanwhile, in order to more quickly approach to the next target point after measurement and search of one target point are completed each time, fitting calculation is carried out once according to the existing target point, and a fitting result is directly used as a gradient descent starting point of the next target point, so that the efficiency of the whole algorithm flow is further accelerated.

Meanwhile, in actual verification, the condition that even if the voltage of a target point with extremely low brightness is 0, the target point with extremely low brightness is still higher than the brightness of the theoretical target point appears, and for the condition, the data of the last target point is directly used to obtain a result through a binding point value interpolation method, so that the integrity of the result is ensured.

In addition, during the gradient descent, several termination conditions are set:

1. when the measured Lv, x, y is sufficiently close to the theoretical target Lv, x, y.

2. When the gradient drops to an upper limit for a certain number of times.

3. When the possible ranges have all been measured.

When the gradient descent is terminated, the combination which is most suitable for the expected combination is found out from the current round of measurement results to be used as the current round of result, and the optimizer is reset to start the approximation of the next target point until all the target points obtain the result.

Drawings

Fig. 1 shows a flow chart of the present invention.

Description of the embodiments

The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Referring to fig. 1, the present invention provides a rapid gamma correction method based on gradient descent.

Step 1, inputting RGB voltage values, measuring the brightness Lv and the color coordinates x and y of the screen body, and comparing the brightness Lv and the color coordinates x and y with target values.

In actual logic, when the measured Lv, x and y are close enough to the theoretical targets Lv, x and y, the measurement of the round is terminated, whether the brightness Lv is close enough or not is judged that two limits act simultaneously, and the judgment that the brightness Lv is close enough can be obtained only if the brightness Lv is satisfied simultaneously, wherein one limit is thatThe value of (2) is less than a predetermined value, such as 0.05nits; at the same time, another percentage limitation that needs to be met, namely +.>Less than a predetermined percentage value, such as 5%, is required. The reason for the design of the double limitation is that a simple numerical threshold may be close to or even smaller than the numerical threshold at low luminance, in which case the result is of course inaccurate; similarly, the deviation of the percentage threshold at high brightness may be several tens of nits, which is also unacceptable. The cost and effect of setting rules to distinguish between high and low brightness are not as good as the two restrictions are used at the same time, so the brightness is used at the same time by using double restrictions. The brightness and color coordinates x, y of the other two measured values only need to meet the condition that the difference between the brightness and color coordinates x, y and the target value is smaller than a threshold value, and the absolute value of the deviation can be ensured because the target values of x and y are fixed in the whole measuring processAccuracy of the results.

In addition, when the gradient descent times reach the threshold value in one round of measurement, the system is also forced to terminate when the gradient descent times still do not converge to meet the limiting condition, and in this case, it is quite possible to indicate that objective factor screens cannot meet the terminating conditions such as close enough no matter what combination is input, and the system is directly terminated in order to avoid excessive meaningless time consumption.

There is also a case where the cycle is terminated after all possible partial combinations are tried or when the voltages of the three RGB are 0 or the number of attempts reaches the upper limit.

If the cycle is not terminated, the step 2 is performed, and if the cycle is terminated, the step 4 is performed.

And 2, calculating the mean square error and the gradient when the comparison result in the step 1 does not meet the termination condition.

The mean square error is calculated asIn the formula +.>Is a target value->I is the sample number, n is the sample group number; taking three groups of measurement samples of brightness Lv and color coordinates x and y of the screen body, wherein n= 3,i =1 corresponds to Lv, i=2 corresponds to x, and i=3 corresponds to y;

the gradient is calculated by the formula:

the gradient provides a direction and a step length, the calculated MSE value is stored in combination with the RGB voltage value for generating the group of data, so as to be used when the optimal result is selected later, and the gradient is used as a parameter to be transmitted to the gradient descent algorithm in the step S3.

And 3, providing the gradient calculated in the step 2 as a parameter to an optimizer, and calculating by the optimizer to obtain the updated numerical value.

Gradient descent is an optimization algorithm that is used to solve for the minimum of the mathematical function. The optimizer is a tool for further optimizing on the basis of the gradient descent algorithm, and can adaptively adjust super parameters such as learning rate and the like according to different requirements, targets and learning models so as to train the optimal model more quickly. In the current task, the brightness along with the voltage change can be regarded as monotonously increasing, so that the minimum value of the loss function is the minimum value, and the local optimal solution is the global optimal solution, so that convergence can be accelerated according to positive and negative gradients with confidence. When the gradient sign is not changed, the learning rate is gradually increased, and the acceleration loss function converges. When the gradient sign is changed, judging the relation between the current learning rate and the initial learning rate, and if the current learning rate is higher than the initial learning rate, directly resetting the relation to the initial learning rate; if the current learning rate is smaller than the initial learning rate, the current learning rate is further reduced when the current learning rate is smaller than the initial learning rate, the current learning rate is further reduced after the oscillation is started and the gradient direction is changed for a plurality of times.

For the optimizer, the following parameters are needed to realize the initial learning rateVariable a, direction d, learning rate->The method comprises the steps of carrying out a first treatment on the surface of the The first two parameters need to be specified at the initialization of the optimizer, i.e. +.>The value of a and a is required to be specified in advance, and the brightness gradient in the gradient calculated in the step 2 is required to be transmitted at each call, and the comparison between the sign of the brightness gradient and the direction d is used for +.>Is updated according to the update of (a); the specific logic is as follows:

wherein,for the current state learning rate, < >>Learning rate for the previous state; finding out learning rate->The post d will be updated to the sign of the brightness gradient and will learn the rate +.>Returning as an optimizer result;

and obtaining the learning rate, updating d into a sign of the brightness gradient after obtaining the learning rate, and returning the learning rate as an optimizer result.

After learning rate is obtained, prediction update can be started to the next voltage combination to be measured.

First, the brightness is adjusted by using the gradient of the brightness LvTo uniformly adjust R, G, B:

wherein,for the last measured RGB voltage value of the current target point +.>，/>，/>；Is a voltage change value, 1 is taken if the voltage change value is between 0 and 1, and-1 is taken if the voltage change value is between-1~0; if the voltage change value has exceeded the input voltage value V, this is taken as +.>；

Secondly, the brightness is adjusted according to the chromaticity after the adjustment, and the rule is to use the gradient of the color coordinate xR is adjusted separately, using the gradient of the color coordinate y +.>The adjustment of G is carried out independently, and is specifically as follows:the voltage is regulated twice in the above sequence, and the regulated voltage is obtained>、、/>After the values are rounded off, the values are transmitted back to the step S1 as input to continue measurement until the termination condition is met, and the optimal result is selected.

And 4, selecting an optimal result when the comparison result in the step 1 meets the termination condition.

The selection of the optimal result is classified into two kinds depending on the method, and the classification is performed depending on which termination condition is satisfied in step 1. If the termination condition is that the deviation threshold is met, the set of voltage combinations meeting the deviation threshold is directly selected as a result. If the triggered termination condition is other conditions, the mean square error MSE of all measurement results in the round of measurement is counted, and a group with the minimum MSE value is selected as the optimal result.

And 5, after finishing the measurement of one target point, judging whether other target points need to be measured, if so, updating the target point, fitting the starting point of the next round, resetting the optimizer, and then starting the measurement of a new target point again. After setting the new RGB voltage input values, go to step 1.

Wherein fitting the starting point of the next round requires the results of the last two rounds as a reference, and skipping this step if there are no results of the last two rounds. The fitting is specifically to calculate the inverse of an exponential term of the results of the previous two rounds, the edge needs to perform fitting calculation on three channels of RGB respectively, binding point values of the previous two rounds are denoted as b1 and b2, voltage values are denoted as v1 and v2, the currently calculated binding point value is denoted as b0, unknown number voltage needed to be fitted is denoted as v0, and a fitting formula of v0 is as follows:

second, for->Performing numerical detection, and if the value is not in the (0, 1) interval, directly using the voltage result of the last target point as a voltage starting value without adopting the fitting result as the voltage starting value of the next target point;

finally, only whenIf the value is not within the (0, 1) interval, the fitting result is not adopted as the next voltage starting value, and the last voltage result is directly used as the voltage starting value. Only when it is greater than 0 and less than 1, the fitting result is considered to be authentic, and the fitting result is used as the voltage initial value.

If there are no other targets, the measurement is ended.

It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.

Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made therein without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (6)

1. A rapid gamma correction method based on gradient descent is characterized in that,

s1, inputting RGB voltage values, measuring brightness Lv and color coordinates x and y of a screen body, and comparing the brightness Lv, the color coordinates x and y with target values of theoretical targets Lv, x and y;

s2, comparing the results obtained in the step S1, and if the termination condition is met, selecting an optimal result; if the termination condition is not met, the RGB voltage value to be measured of the current target point is adjusted, and the step S1 is repeated until the termination condition is met;

each time the RGB voltage value to be measured last time is adjusted, the RGB voltage value combination to be measured next time is updated by using a gradient descent algorithm, and the combination value is used as input to return to the step S1 for continuous measurement;

s3, adjusting RGB voltage values to be measured of the next target point to obtain an optimal result; if other target points exist, respectively acquiring an optimal result of each target point, namely repeating the steps S1-S2;

s4, if each target point is measured, an optimal result is obtained, and the measurement is finished;

the adjusting of the RGB voltage value to be measured of the current target point comprises:

mean square error calculation, the formula isIn the formula +.>Is a target value->I is the sample number, n is the sample group number; taking three groups of measurement samples of brightness Lv and color coordinates x and y of the screen body, wherein n= 3,i =1 corresponds to Lv, i=2 corresponds to x, and i=3 corresponds to y;

the gradient is calculated by the formula:the gradient provides a direction and a step length, the calculated MSE value is combined with the RGB voltage value for generating the group of data and stored for later use when the optimal result is selected, and the gradient is used as a parameter to be transmitted to a gradient descent algorithm in the step S2;

in step S2, providing the gradient calculated in step S2 as a parameter to an optimizer, and calculating the value updated at this time by the optimizer;

the brightness Lv increases monotonically with the voltage change, the minimum value of the loss function is the minimum value, and the local optimal solution is the global optimal solution, so that convergence is accelerated according to positive and negative of the gradient:

s2.1, gradually increasing the learning rate when the gradient sign is not changed, and accelerating the convergence of the loss function;

s2.2, when the gradient sign changes, judging the relation between the current learning rate and the initial learning rate:

(1) If the current learning rate is higher than the initial learning rate, directly resetting to the initial learning rate;

(2) If the current learning rate is smaller than the initial learning rate, the current learning rate is further reduced when the current learning rate is smaller than the initial learning rate, the current learning rate is further reduced after the oscillation is started and the gradient direction is changed for a plurality of times;

the optimizer is realized by the following parameters: initial learning rateVariable a, direction d, learning rate->；

The first two parameters, namely the initial learning rate, need to be specified when the optimizer is initializedAnd the value of the variation parameter a, which is required to be specified in advance, the luminance gradient +_ among the gradients calculated in step S2 needs to be entered at each call>，Wherein->For the measurement of the brightness Lv +.>For the target value of the luminance Lv, the learning rate +.>Is updated according to the update of (a);

the learning rateThe updating steps of (a) are as follows: /> Wherein (1)>For the current state learning rate, < >>Learning rate for the previous state; finding out learning rate->The post d will be updated to the sign of the brightness gradient and will learn the rate +.>Returning as an optimizer result;

obtaining the learning rateThen, the prediction updating is carried out on the voltage combination to be measured next time;

first, according to brightness adjustment, a rule is to use gradient of brightness Lv simultaneouslyTo uniformly adjust R, G, B:

wherein (1)>For the last measured RGB voltage value of the current target point +.>，/>，/>；/>Is a voltage change value, 1 is taken if the voltage change value is between 0 and 1, and-1 is taken if the voltage change value is between-1~0; if the voltage change value has exceeded the input voltage value V, this is taken as +.>；

Secondly, the brightness is adjusted according to the chromaticity after the adjustment, and the rule is to use the gradient of the color coordinate xR is adjusted separately, using the gradient of the color coordinate y +.>The adjustment of G is carried out independently, and is specifically as follows: />The voltage is regulated twice in the above sequence, and the regulated voltage is obtained>、/>、/>After the values are rounded off, the values are transmitted back to the step S1 as input to continue measurement until the termination condition is met, and the optimal result is selected.

2. The rapid gamma correction method based on gradient descent according to claim 1, wherein the termination condition judgment method in step S2 is as follows:

when the measured values of the brightness Lv and the color coordinates x, y of the measurement screen are close enough to the target values of the theoretical targets Lv, x, yTerminating the measurement of the present round, determining whether the luminance Lv satisfies sufficiently close requirements while satisfying "two limitations", one of which isThe value of (2) is smaller than the first threshold, the other limitation is the percentage limitation, i.e. +.>Less than a second threshold.

3. The gradient descent-based rapid gamma correction method as claimed in claim 2, wherein it is determined whether the other two measured color coordinates x, y are sufficiently close to each other to satisfyA value less than a third threshold value, ">The value is less than the fourth threshold.

4. The rapid gamma correction method according to claim 1, wherein,

if the termination condition satisfied in step S1 includes at least one of the following conditions:

(1) When the gradient descends to reach the upper limit of a certain number of times;

(2) When the possible ranges have all been measured;

the mean square error MSE of all the measurement results in this round of measurement is counted and a group with the smallest MSE value is selected as the optimal result.

5. The rapid gamma correction method according to claim 1, wherein if there are other target points in step S3, the following steps are performed:

s3.1, updating a target point;

s3.2, fitting the starting point of the next round;

s3.3, resetting the optimizer;

s3.4, starting measurement of a new target point again;

wherein S3.2 fitting the starting point of the next round requires the results of the last two rounds as a reference, and if there are no results of the last two rounds this step S3.2 is skipped.

6. The gradient descent-based rapid gamma correction method of claim 5, wherein the fitting is specifically performed by calculating the inverse of the exponential term of the previous two rounds of results:

firstly, fitting calculation needs to be carried out on three RGB channels, binding point values of the previous two rounds are denoted as b1 and b2, voltage values are denoted as v1 and v2, currently calculated binding point values are denoted as b0, unknown number voltages needing fitting are denoted as v0, and a fitting formula of v0 is as follows:

second, for->Performing numerical detection, and if the value is not in the (0, 1) interval, directly using the voltage result of the last target point as a voltage starting value without adopting the fitting result as the voltage starting value of the next target point;

finally, only whenAbove 0 and below 1, the fitting result is considered to be authentic, using the fitting result v0 as the initial voltage value.