KR102527718B1 - Method for optimization of lens module assembly - Google Patents

Abstract

In the lens module assembly optimization method according to an embodiment of the present invention, in assembling a lens module in which N lenses generated in corresponding cavities are assembled to overlap along an optical axis, a computing system performs a process between N lenses, respectively. receiving characteristic information of at least N lenses formed in N cavity groups including a plurality of cavities having a higher degree of identity than each other; and processing, by the computing system, information for selecting N cavities from the N cavity groups based on the characteristic information; The computing system includes a past cavity selection result by processing the selection information, a fitness function configured based on data of N lenses or assembled lens modules according to the past cavity selection result, and genetic In the step of receiving or storing the algorithm and processing the selection information, the chromosome individual information is obtained based on the fitness function and the crossover or mutated output chromosome information based on the genetic algorithm from the input chromosome information corresponding to the past cavity selection result. updating, and processing information for selecting N cavities based on chromosome entity information and characteristic information.

Description

Lens module assembly optimization method {Method for optimization of lens module assembly}

The present invention relates to a lens module assembly optimization method.

Recently, as a number of high-performance camera modules are installed in smartphones, camera module manufacturers are focusing on maximizing production efficiency to respond to rapid demand growth using limited factories and facilities. The camera module may include a lens module in which a plurality of lenses are compactly arranged, and mass productivity of the lens module is becoming increasingly difficult to secure as the camera module increases in performance.

Currently, in order to find assembly conditions for mass production of lens modules, field workers select assembly conditions that show high performance after manually trying a number of assembly conditions based on the operator's knowledge and past experience. This has a disadvantage in that it consumes time and money and cannot even try for conditions unknown to the operator.

Registered Patent Publication No. 10-1512829

The present invention provides a method for optimizing the assembly of a lens module in which N lenses generated in corresponding cavities are assembled to overlap along an optical axis.

In the lens module assembly optimization method according to an embodiment of the present invention, in assembling a lens module in which N (N is a natural number of 2 or more) lenses generated in corresponding cavities are assembled to overlap along an optical axis. , Receiving, by a computing system, characteristic information of at least N lenses formed in N cavity groups each including a plurality of cavities each having a higher degree of identity than between the N lenses; and processing, by the computing system, information for selecting N cavities from the N cavity groups based on the characteristic information. The computing system includes a fitness function configured based on past cavity selection results obtained by processing the selection information and data of N lenses or assembled lens modules according to the past cavity selection results. In the step of receiving or storing a genetic algorithm and processing the selection information, output chromosome information crossed or mutated based on the genetic algorithm from input chromosome information corresponding to the past cavity selection result and the fitness function It may be characterized in that chromosome entity information is updated based on , and information for selecting N cavities is processed based on the chromosome entity information and the characteristic information.

In the lens module assembly optimization method according to an embodiment of the present invention, in assembling a lens module in which N (N is a natural number of 2 or more) lenses generated in corresponding cavities are assembled to overlap along an optical axis. , Receiving, by a computing system, characteristic information of at least N lenses generated from N cavity groups including a plurality of cavities each having a higher degree of identity than between the N lenses; processing, by the computing system, information for selecting N cavities from the N cavity groups based on a fitness function and the characteristic information; and rotation of at least one optical axis of the N lenses corresponding to the N cavities when assembled so that the N lenses corresponding to the N cavities overlap along the optical axis, based on the selection information. processing the angle information to be; The computing system includes a past cavity selection result by processing the selection information and the fitness degree configured based on data of N lenses or assembled lens modules according to the past cavity selection result. A function is provided or stored, and the computing system updates the fitness function based on data of each of the N lenses corresponding to the N cavities selected by the processing of the selection information, and processes the angle information. It may be characterized in that the update of the fitness function based on the data of the lens module assembled at the angle corresponding to the step is performed according to a machine learning algorithm.

The lens module assembly optimization method according to an embodiment of the present invention can reduce trial and error in the process of assembling the lens module or an operator who assembles the lens module, and can efficiently improve the mass productivity of the lens module.

1A is a diagram showing that a lens module assembly optimization method according to an embodiment of the present invention is applied to a lens module assembly process.
1B is a diagram illustrating a cavity that may be used to form a lens.
2A and 2B are flowcharts illustrating a lens module assembly optimization method according to an embodiment of the present invention.
3 is a block diagram showing a cavity selection structure of a lens module assembly optimization method according to an embodiment of the present invention.
4 is a diagram illustrating chromosomal information of a genetic algorithm of a lens module assembly optimization method according to an embodiment of the present invention.
5 and 6 are diagrams illustrating chromosome mating of the genetic algorithm of the lens module assembly optimization method according to an embodiment of the present invention.
7 is a block diagram showing a rotation angle prediction structure when assembling a lens in a lens module assembly optimization method according to an embodiment of the present invention.
8 is a diagram illustrating an interaction between a lens module assembly optimization method and a provider according to an embodiment of the present invention.
FIG. 9 is a diagram illustrating a graphical user interface (GUI) of FIG. 8 .

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The detailed description of the present invention which follows refers to the accompanying drawings which illustrate, by way of illustration, specific embodiments in which the present invention may be practiced. These embodiments are described in sufficient detail to enable one skilled in the art to practice the present invention. It should be understood that the various embodiments of the present invention are different from each other but are not necessarily mutually exclusive. For example, specific shapes, structures, and characteristics described herein may be implemented in one embodiment in another embodiment without departing from the spirit and scope of the invention. Additionally, it should be understood that the location or arrangement of individual components within each disclosed embodiment may be changed without departing from the spirit and scope of the invention. Accordingly, the detailed description set forth below is not to be taken in a limiting sense, and the scope of the present invention is limited only by the appended claims, along with all equivalents as claimed by those claims. Like reference numbers in the drawings indicate the same or similar function throughout the various aspects.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings so that those skilled in the art can easily practice the present invention.

In addition, parts represented by the same color in a plurality of drawings of the present specification may be implemented identically or similarly to each other.

FIG. 1A is a diagram showing that a lens module assembly optimization method according to an embodiment of the present invention is applied to a lens module assembly process, and FIG. 1B is a diagram illustrating a cavity that can be used to form a lens.

Referring to FIG. 1A , the lens module 40 may be assembled such that N lenses 20 overlap along an optical axis (eg, a vertical direction). For example, a combined structure 30 in which N lenses 20 and a plurality of spacers 12 are alternately arranged may be assembled into the lens barrel 13 . N is not limited to 7 and may be a natural number of 2 or more.

The N lenses 20 may be designed to have different optical or physical parameters (eg, diameter, thickness, curvature), and the characteristics of the lens module 40 may be a combination of the N lenses 20 having different characteristics. may vary as much as the number of cases of

Each of the N lenses 20 may be formed in a corresponding cavity among the N cavities. For example, each of the N lenses 20 may be produced by injection molding as a resin of a transparent or lightweight material (eg, plastic) is injected into a cavity and cured in a high-temperature atmosphere.

As the number of N cavities corresponding to the N lenses 20 increases, the number of lens modules 40 assembled per unit time may increase, and mass productivity of the lens modules 40 may further increase.

Accordingly, the N cavities corresponding to the N lenses 20 may be composed of the N cavity groups 11-1, 11-2, 11-3, and 11-N, and the N cavity groups 11-1 , 11-2, 11-3, 11-N) may each include a plurality of cavities.

A plurality of cavities of each of the N cavity groups 11 - 1 , 11 - 2 , 11 - 3 , and 11 -N may have a higher identity to each other than between the N lenses 20 . A plurality of lenses of each of the N lenses 20 may be simultaneously formed in a plurality of corresponding cavities.

Referring to FIG. 1B , the first cavity group 11-1 may include M first cavities C ₁ , C ₂ , C _m , and M first cavities C ₁ , C ₂ , C _m ) may be the same as each other. M is not limited to 16 and may be a natural number of 2 or more. In addition, the number of cavities included in each of the N cavity groups may be the same as 16, but is not limited thereto.

For example, the M first cavities (C ₁ , C ₂ , C _m ) may be connected to each other through connecting members (B ₁ , B ₈ ), and the core members (A 1 , A 2 ) may be connected to each other (B ₁ , A ₂ ). ₁ , B ₈ ), and may be configured to allow a lens assembly process or a person to hold the first cavity group 11-1.

For example, the M number of first cavities (C ₁ , C ₂ , C _m ) included in the first cavity group 11-1 are formed by connecting members (B ₁ , B ₈ ) before the M number of first lenses are formed. After the M number of first lenses are formed in a state connected to each other through, they may be separated from each other. A portion where the connecting members B ₁ and B ₈ are connected in each of the M first cavities C ₁ , C ₂ , and C _m may be used as a reference point for a rotation angle when assembling the lens.

Since the characteristics of the N lenses 20 may depend on the characteristics of the cavities, the cavities may have characteristics dependent on the design of optical or physical parameters of each of the N lenses 20 . Since the optical or physical parameters of the N lenses 20 may require high delicacy, a high correlation between the design of the N lenses 20 and the characteristics of the cavity may be required. In the process of creating N lenses 20 in the cavity, various variables (eg, lens clamping force, resin filling time, temperature, cooling rate, etc.) act as limitations in pursuing the degree of correlation. It can be.

Since various variables in the process of generating the N number of lenses 20 in the cavity may act minutely differently between the first cavities C ₁ , C ₂ , and C _m of the first cavity group 11-1, the M number of Optical or physical parameters of the M number of first lenses generated in one cavity (C ₁ , C ₂ , C _m ) may be slightly different from each other.

Referring back to FIG. 1A , optical or physical parameters of M lenses formed in each of the M cavities of the remaining cavity groups 11-2, 11-3, and 11-N other than the first cavity group 11-1 may also differ slightly from each other.

From the first difference between the M number of first lenses to the Nth difference between the M number of Nth lenses, a secondary effect may be produced according to their combination. The negative effect of the secondary effect on the characteristics of the assembled lens module 40 may vary as much as the number of cases of the secondary effect, and the number of cases of the secondary effect (eg, M ^N ) may be large. there is. The lens module data 102 may include the secondary effect.

For example, the secondary effect may cause an overlapping phenomenon due to aberration and diffraction in the process of at least a portion of light passing through the assembled lens module, and the degree of light overlapping is defined as MTF (Modulation Transfer Function) It can be. The MTF may be defined as a ratio between an overlapping modulation value of an object and an overlapping modulation value of an upper surface of a lens module, and an ideal MTF may be 1. If the MTF is out of the standard range, the assembled lens module may be determined to be defective.

The computing system 100a performing the lens module assembly optimization method according to an embodiment of the present invention reduces the secondary effect of the lens module 40 to be assembled before the N lenses 20 are assembled, or the lens module ( 40), the cavity assembly conditions 107 corresponding to the information for selecting N cavities can be efficiently provided so as to reduce the defect rate.

Depending on the design, the computing system 100a reduces the secondary effect of the lens module 40 to be assembled or the defective rate of the lens module 40 based on the rotation angle assembly condition 109 according to the cavity assembly condition 107. this can be reduced The rotation angle assembling condition 109 may include information on an angle rotated in a direction of winding at least one optical axis among the N lenses when the N lenses corresponding to the N cavities are assembled to overlap along the optical axis.

For example, the isotropic of the generated N lenses 20 may be lower than that of the corresponding cavity due to various variables in the process in which the N lenses 20 are created in the cavity, and the generated N lenses 20 The roundness or flatness of (20) may deviate further by 1. An influence of the anisotropic elements of the N lenses 20 on the secondary effect may be determined based on the rotation angle assembly condition 109 .

For example, the recording medium 99 is configured to be accessible to the computing system 100a in order to provide recorded information to the computing system 100a so that the computing system 100a can perform the lens module assembly optimization method. It can be.

2A and 2B are flowcharts illustrating a lens module assembly optimization method according to an embodiment of the present invention.

Referring to FIG. 2A , the computing system performing the lens module assembly optimization method according to an embodiment of the present invention, in a state in which (N*M) lenses are formed (S20), each plurality (eg, M lenses) Characteristic information of at least N lenses formed in N cavity groups including cavities of may be provided (S101), and a fitness function may be provided (S124), and based on the characteristic information, N lenses may be provided. Information for selecting N cavities from the cavity group may be processed (S120), and a combination selected from among M ^N cases of cavity combinations may be generated (S107).

A computing system performing a lens module assembly optimization method according to an embodiment of the present invention efficiently reduces or compresses a large number of cavity combinations based on a genetic algorithm, and provides among the reduced number of cavity combinations. A combination of cavities can be selected according to the characteristic information to be used. Accordingly, trial and error occurring in the process of assembling the lens module by an operator or an assembly process may be reduced.

Referring to FIG. 2B , in the lens module assembly optimization method according to an embodiment of the present invention, a computing system performing the lens module assembly optimization method according to an embodiment of the present invention forms (N*M) lenses. In the state of (S20), characteristic information of at least N lenses formed in N cavity groups each including a plurality of (eg, M) cavities may be provided (S101), based on the fitness function and the characteristic information. The information for selecting N cavities from the N cavity groups can be processed (S120), rotation angle information can be processed based on the selected information (S125), and the fitness function can be updated (S126) there is.

The computing system performing the lens module assembly optimization method according to an embodiment of the present invention, in the step of processing the selection information (S120), the fitness function based on the data of each of the N lenses corresponding to the N cavities selected. Updating and updating the fitness function based on the data of the lens module assembled at the angle corresponding to the step of processing the angle information (S125) may be performed according to a machine learning algorithm.

The information for selecting the cavity and the angle information may have a time-series and mutually influencing relationship, and the machine learning algorithm may be efficient in processing data having a time-series and mutually influencing relationship. Therefore, the goodness-of-fit function can be efficiently updated, and a great contribution can be made to the computing system to select an efficiently selected cavity combination among a number of cases of cavity combinations.

3 is a block diagram showing a cavity selection structure of a lens module assembly optimization method according to an embodiment of the present invention, and FIG. 4 illustrates chromosome information of a genetic algorithm of a lens module assembly optimization method according to an embodiment of the present invention. it is a drawing

Referring to FIG. 4, chromosome information 111 used in the genetic algorithm 110b includes N or M point information 113-1, 113-2, 113-3, 113-4, 113-5, 113-6, 113-N) and M or N gene information (112-1, 112-2, 112-M).

For example, since a combination of (N*M) cavities corresponding to (N*M) lenses can be divided into M cavity combinations each including N cavities, one chromosome information 111 is M It may include the gene information (112-1, 112-2, 112-M), M number of lens modules may be assembled.

Referring to FIG. 3 , the computing system 100a-1 performing the lens module assembly optimization method according to an embodiment of the present invention initializes 121 to generate chromosome information corresponding to each of the M lens modules, The generated chromosomal information is evaluated (122) based on the fitness function 120, and excellent genetic information is selected from among M genetic information or excellent chromosome information is selected from among a plurality of chromosome information based on the evaluation result and the genetic algorithm (110a). selection (110-1), crossover (110-2) and/or mutation (110-3) of the input chromosome information based on the selection result to generate output chromosome information, and evaluation of the output chromosome information (123); Based on the evaluation result, the chromosome entity information 130 may be updated. These processes can be repeated.

In general, the computing system 100a-1 outputs crossover or mutated output chromosome information based on the genetic algorithm 110a from input chromosome information corresponding to past cavity selection results and chromosome individual information based on the fitness function 120. 130 may be updated, and cavity assembly conditions 107 including information for selecting N cavities based on the chromosome entity information 130 and the lens characteristic information 101 may be generated.

Accordingly, a large number of combinations of cavities can be efficiently reduced or compressed based on a genetic algorithm.

For example, the update of the chromosome entity information 130 depends on the number of values greater than or equal to a reference value among a plurality of values according to the positive/non-prediction model 124 of the plurality of gene information included in the input chromosome information and/or the output chromosome information. It may be performed based on the number of values that are equal to or greater than the reference value among a plurality of values according to the pass/fail prediction model 124 of the selection results of the M cavities selected from the N cavity groups.

Since the computing system 100a-1 may be provided with the lens characteristic information 101, a corresponding relationship between the update of the chromosome entity information 130 and the lens characteristic information 101 may be known, and the lens characteristic information ( 101), specific chromosome information may be selected from the chromosome entity information 130, and the selected chromosome information may correspond to information for selecting N cavities. For example, the lens characteristic information 101 may be optical or physical measurement information (eg, diameter, thickness, curvature, roundness, flatness) of N lens groups.

The computing system 100a-1 may repeatedly learn (100a-3) the pass/fail prediction model 124 based on the N lenses or lens module data 102 according to the past cavity selection results, and The function 120 can be updated iteratively.

The fitness function is calculated by Equation 1 below, c_i means information on the i-th chromosome, and n(·) may represent a function for calculating the number of gene information determined to be good products among chromosome information. For example, genetic information may be determined as good or bad based on the good/non-good prediction model 124 .

[Equation 1]

In this case, the pre-learned pass/fail prediction model 124 is used to determine the quality of the gene, and the pass/fail prediction model 124 may use a methodology suitable for the user's purpose. The predictive model can directly predict the value of the lens module (eg, MTF) using the regression model, and the pass/fail classification methodology of the example below can be converted to the regression model as it is.

For example, a possible methodology with the pass/fail prediction model 124 is as follows. As a general classification methodology, the following methodologies can be applied.

A support vector machine (SVM) is a model that finds the hyperplane that best divides a class when solving a classification problem that divides classes. The hyperplane means the point that maximizes the margin(w), the distance between the plus-plane and the minus-plane, and the margin is determined by the position of the hyperplane and the position of the support vector. SVM has the advantage of having no influence on outliers, but has the disadvantage of taking a relatively long time to build a model and having low explanatory power for the results.

A decision tree (DT) is a classification model that divides the independent variable space while sequentially applying various rules. It has the advantage of low computational cost and high explanatory power of the results.

A random forest (RF) is an ensemble technique that utilizes bootstrap aggregating or bagging to partition training and test data into multiple subsets. Bagging is a basic technique of RF and has the effect of reducing the variance of predicted values. With higher accuracy, it is possible to handle thousands of input variables.

Bayes classifier is a classification technique that predicts the class that best represents data by solving a stochastic problem based on Bayesian probability. In general, it is a simple calculation that simplifies the calculation by assuming the conditional independence of each class label when calculating the posterior probability. The Naive Bayes classification model is mainly used. In addition, the amount of training data for estimating the parameters is very small, and it is robust against outliers.

Furthermore, various methodologies can be applied to compensate for the data imbalance problem that occurs because good product data is mainly used at manufacturing sites, but bad data is rarely collected.

Balanced bagging classifier is one of the ensembles specialized for class imbalanced data, which aggregates after predicting arbitrary subsets obtained from a given data set with a basic classifier.

RUSBoost classifier and SMOTEBoost classifier are imbalanced data classification methods that hybridize boosting concept with undersampling and oversampling techniques, respectively. They eliminate data variance imbalance and improve the performance of poor classifiers.

MFE (mean false error) based DNN (deep neural network) is a method that uses MFE or MSFE (mean squared false error) proposed as a loss function by focusing on data imbalance problems in general DNN methodologies. In general, MSE (mean squared error) is widely used as a loss function, which represents the average of error values for all instances and is considered the same regardless of class. However, MFE calculates the average error value considering the total number of instances per class. For this, in Equation 2, mean false error, false positive error, and false negative error are respectively shown.

[Equation 2]

5 and 6 are diagrams illustrating chromosome mating of the genetic algorithm of the lens module assembly optimization method according to an embodiment of the present invention.

Referring to FIG. 5, a plurality of pieces of input chromosome information 111-11 and 111-12 used in the genetic algorithm 110c may correspond to parent chromosomes, and output chromosome information 111-21 may correspond to child chromosomes. It can be.

The output chromosome information 111-21 may be generated through point crossover between the plurality of input chromosome information 111-11 and 111-12.

For example, the first to fourth point information 113-41 of the first input chromosome information 111-11 is combined with the first to fourth point information 113-41 of the output chromosome information 111-21. The fifth to seventh point information 113-52 of the second input chromosome information 111-12 is the fifth to seventh point information 113-52 of the output chromosome information 111-21. can be the same as The location of the point to be cross-blended can be determined randomly.

Referring to FIG. 6, a plurality of pieces of input chromosome information 111-13 and 111-14 used in the genetic algorithm 110d may correspond to parent chromosomes, and output chromosome information 111-22 may correspond to child chromosomes. It can be.

The output chromosome information 111-21 may be generated through gene crossover between a plurality of input chromosome information 111-13 and 111-14.

For example, among the M pieces of gene information of the second input chromosome information 111-14, the second gene information 112-22 has the highest probability of defective prediction according to the pass/fail prediction model or the lens predicted by the regression model. The numerical value (eg, MTF) of the module may be the lowest gene information with a reference value or less, and may correspond to a recessive gene. Among the M pieces of gene information of the first input chromosome information (111-13), the second gene information (112-21) has the highest probability of predicting a good product according to the pass/fail prediction model or the MTF value predicted by the regression model is greater than or equal to the reference value. , and may be the highest genetic information, and may correspond to a dominant gene. The second gene information 112-22 of the second input chromosome information 111-14 is the second gene information of the first input chromosome information 111-13 in the copy 111-15 of the second input chromosome information 112-21).

In addition, specific point information of the second gene information 112-23 of the output chromosome information 111-21 may be mutated. The position of the point to be displaced may be randomly determined.

7 is a block diagram showing a rotation angle prediction structure when assembling a lens in a lens module assembly optimization method according to an embodiment of the present invention.

Referring to FIG. 7 , the computing system 100a-2 performing the lens module assembly optimization method according to an embodiment of the present invention includes a cavity assembly condition 107 including information for selecting a cavity and a rotation angle prediction model. Based on (126), the rotation angle assembly condition can be predicted (125), the rotation angle assembly condition (109) is generated, and the cavity assembly condition (108) in which the variables of the rotation angle assembly condition (109) are further reflected can be generated. can

The data 102 of the lens module assembled according to the cavity assembly condition 108 in which the variables of the rotation angle assembly condition 109 are further reflected and the angle information corresponding to the rotation angle assembly condition 109 are model learning (100a-4) , and the learning target of the model learning (100a-4) may be the rotation angle prediction model 126, and may be a pass/fail prediction model and/or a goodness-of-fit function shown in FIG. 3.

That is, the pass/fail prediction model and/or the fit function may be updated based on data (eg, characteristic information) of each of the N lenses corresponding to the N cavities as shown in FIG. 3, and the rotation angle assembly It may be updated based on the data 102 of the lens module assembled according to angle information corresponding to the cavity assembly condition 108 and the rotation angle assembly condition 109 in which the variables of the condition 109 are further reflected. The rotation angle prediction model 126 may be performed according to an efficient machine learning algorithm (eg, RNN) for processing data having time-series and mutually influencing relationships. For example, the rotation angle prediction model 126 may receive characteristic information of N lenses and generate N angle information.

The rotation angle prediction model 126 inputs the lens module production data into the prediction model, predicts N rotation angle combinations corresponding to each dataset, and learns so that the difference from the actual N rotation angle combinations converges to zero. It can be.

For example, the methodology possible with the machine learning algorithm and the rotation angle prediction model 126 is as follows.

LSTM (long short term memory) (Hochreiter and Schmidhuber, 1997) is one of the recurrent neural network (RNN)-based methodologies known for its excellent performance in processing time series data. It is used for prediction by putting more weight on information that needs to be separated and remembered.

In the RNN structure, a cell state is added in addition to the hidden layer, and three newly added gates are used to obtain the hidden state value and the cell state value.

The three added gates are (1) forget, (2) input, and (3) output gates, respectively, and the activation function expression for each gate is as shown in Equation 3-5.

[Equation 3]

[Equation 4]

[Equation 5]

The Forget gate is used to control the hidden layer nodes where past information is stored. f _t is calculated as a value of 0 or 1. If it is 0, the past information is removed and not used, and if it is 1, the past information keep

Input gates are used to determine what information needs to be retained in order to update the current cell state.

i _t is also calculated as a value of 0 or 1 like f _t . If it is 0, the past information is maintained, and if it is 1, it is updated using information inherent in the current cell state.

Finally, in the output gate, it is used to determine what information to display as output, and is calculated as a value of 0 or 1. In case of 0, it is not displayed and in case of 1, it is displayed.

C-LSTM (convolutional long short term memory) (Shi et al., 2017) is a methodology that combines deep convolutional neural networks (DCNN) (Krizhevsky et al., 2012) and bidirectional LSTM (Graves et al., 2009).

DCNN is a methodology known to effectively capture features inherent in data, and extract feature vectors inherent in input data through a convolutional layer.

Then, based on the introduced unidirectional LSTM or bidirectional LSTM, the output value is predicted by reflecting sequential characteristics.

A gated recurrent unit (GRU) (Cho et al., 2014), like LSTM, is one of the representative cell types of recurrent neural networks.

It has a cell structure with reduced computational complexity while maintaining the advantages of LSTM, and there are only two gates, an update gate and a reset gate.

Both gates reflect the current input value (x_t) and the previous hidden state value (h_(t-1)).

Equation 6-7 below shows the update gate and reset gate calculation in order.

[Equation 6]

[Equation 7]

Equation 8 represents information worth remembering at the current point in time, t, and reflects the current point in time information (Wx_t) and the previous information (Uh_(t-1)), but how much the past information is reflected depends on the reset gate value. have

[Equation 8]

In addition, the above methodologies can be developed through the following mechanisms.

The attention mechanism (Bahdanau et al., 2014) is designed to compensate for the disadvantage that the RNN-based seq2seq model must always be represented as a vector with the same length regardless of the length of the existing input data. has been suggested

Select vectors associated with the target data to be predicted from the input data to create subsets, and use them to predict the target data.

A method of utilizing input data that has an influence on the target data to be predicted at that time with high attention while referring to the entire input data once again at every time-step that outputs the target data through this method am.

8 is a diagram illustrating an interaction between a lens module assembly optimization method and a provider according to an embodiment of the present invention.

Referring to FIG. 8 , a computing system 100b performing a lens module assembly optimization method according to an embodiment of the present invention may include a graphical user interface (GUI) 140, a database 150, and a processor 160. can

For example, computing system 100b may be a personal computer, server computer, handheld or laptop device, mobile device (mobile phone, PDA, media player, etc.), multiprocessor system, consumer electronic device, mini computer, mainframe computer, distributed computing environments that include any of the foregoing systems or devices; and the like. For example, computing system 100b may further include communication connections such as a modem, network interface card (NIC), integrated network interface, radio frequency transmitter/receiver, infrared port, USB connection.

The provider 50 may be a worker who assembles the lens module or may be a lens module assembly process, inputs lens information (eg, the number of lenses, design information) into the graphic user interface 140 (S141), and parameters (eg: Lens characteristic information, actual measurement information) may be input into the graphic user interface 140 (S142), a save button may be clicked (S143), an execution button may be clicked (S144), and assembly optimization results may be displayed (S145) can be checked.

The graphic user interface 140 may provide an environment for interaction between the provider 50 and the computing system 100b, may retrieve lens information from the database 150 (S151), and transmit an assembly optimization result. It can be received (S162).

The database 150 may store information about a fitness function, information about a genetic algorithm, and information about a machine learning algorithm, and transmit lens data to the processor 160 (S152). For example, the database 150 may be implemented with volatile memory (eg, RAM, etc.), non-volatile memory (eg, ROM, flash memory, etc.), or a combination thereof, and includes magnetic storage, optical storage, and the like. It can also be implemented with the same additional storage.

The processor 160 may process information about a fitness function, information about a genetic algorithm, and information about a machine learning algorithm. For example, the processor 160 may include a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor, an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Arrays (FPGA), and the like, , may have a plurality of cores.

FIG. 9 is a diagram illustrating a graphical user interface (GUI) of FIG. 8 .

Referring to FIG. 9 , the graphic user interface 140 includes an area 141 for inputting lens information and parameters, an area 143 with program operation buttons, and an area 145 displaying assembly optimization results. can include

For example, graphical user interface 140 may include input devices (eg, keyboard, mouse, pen, voice input device, touch input device, infrared camera, video input device) and output devices (eg, display, speaker, printer). can include

In the above, the present invention has been described by specific details such as specific components and limited embodiments and drawings, but these are provided to help a more general understanding of the present invention, and the present invention is not limited to the above embodiments. , Those skilled in the art to which the present invention pertains may seek various modifications and variations from these descriptions.

C ₁ , C ₂ , C _m : Multiple (or M) cavities
11-1, 11-N: Cavity group
20: N lenses
100a: Computing system for performing lens module assembly optimization method
110a: genetic algorithm
111: chromosome information
112-1, 112-M: multiple gene information
113-1, 113-N: N point information

Claims (16)

In the assembly of the lens module in which N lenses (N is a natural number of 2 or more) generated in each corresponding cavity are assembled to overlap along the optical axis,
receiving, by a computing system, characteristic information of at least N lenses formed in N cavity groups each including a plurality of cavities each having a higher degree of identity than between the N lenses; and
processing, by the computing system, information for selecting N cavities from the N cavity groups based on the characteristic information; including,
The computing system includes a past cavity selection result by processing the selection information, a fitness function configured based on data of N lenses or assembled lens modules according to the past cavity selection result, and a genetic algorithm. receive or store;
In the step of processing the selection information, chromosome individual information is updated based on the fitness function and output chromosome information crossed or mutated based on the genetic algorithm from the input chromosome information corresponding to the past cavity selection result, A lens module assembly optimization method characterized in that processing information for selecting N cavities based on the chromosome entity information and the characteristic information.
According to claim 1,
The chromosomal information used in the genetic algorithm is composed of a plurality of gene information each having N points (point) information, or composed of N gene information each having a plurality of point information Lens module assembly optimization method.
According to claim 2,
The output chromosome information is generated through at least one of point crossover and gene crossover between a plurality of input chromosome information.
According to claim 2,
Lens module assembly optimization method in which the update of the chromosome individual information is performed based on the number of values greater than or equal to a reference value among a plurality of values according to a pass/fail prediction model of a plurality of gene information included in the input chromosome information or the output chromosome information .
According to claim 1,
Each of the N cavity groups includes M cavities (M is a natural number of 2 or more),
The lens module assembly optimization method in which the update of the chromosome entity information is performed based on the number of values greater than or equal to a reference value among a plurality of values according to a pass/fail prediction model of the selection results of the M cavities selected from the N cavity groups.
According to claim 1,
A lens module assembly optimization method in which a plurality of lenses corresponding to a plurality of cavities included in one of the N cavity groups are formed at the same time.
According to claim 1,
The characteristic information of the N lens groups includes optical or physical measurement information of the N lens groups.
According to claim 1,
The fit function is repeatedly updated based on data of N lenses or assembled lens modules according to the past cavity selection results.
According to claim 1,
Based on the selection information, the computing system is rotated in a direction of winding at least one of the optical axes among the N lenses when the N lenses corresponding to the N cavities are assembled to overlap along the optical axis. A lens module assembly optimization method further comprising processing angle information.
According to claim 9,
The update of the fitness function based on the data of each of the N lenses corresponding to the N cavities selected by the step of processing the selected information, and the data of the lens module assembled at the angle corresponding to the step of processing the angle information Lens module assembly optimization method further comprising performing an update of the fitness function based on a machine learning algorithm.
In the assembly of the lens module in which N lenses (N is a natural number of 2 or more) generated in each corresponding cavity are assembled to overlap along the optical axis,
receiving, by a computing system, characteristic information of at least N lenses generated from N cavity groups including a plurality of cavities each having a higher degree of identity than between the N lenses;
processing, by the computing system, information for selecting N cavities from the N cavity groups based on a fitness function and the characteristic information; and
Based on the selection information, the computing system is rotated in a direction of winding at least one of the optical axes among the N lenses when the N lenses corresponding to the N cavities are assembled to overlap along the optical axis. processing angle information; including,
The computing system receives the fitness function configured based on past cavity selection results obtained by processing the selection information and data of N lenses or assembled lens modules according to the past cavity selection results, or save,
The computing system updates the fitness function based on data of each of the N lenses corresponding to the N cavities selected by the processing of the selection information, and assembles the angle corresponding to the processing of the angle information. A lens module assembly optimization method, characterized in that performing the update of the fitness function based on the data of the lens module according to a machine learning algorithm.
According to claim 11,
The processing of the angle information may include processing the angle information by applying the selected information to a rotation angle prediction model updated based on data of a lens module assembled at an angle corresponding to the processing of the angle information. Module assembly optimization method.
According to claim 11,
Each of the N cavity groups includes M cavities (M is a natural number of 2 or more),
The fitness information output from the fitness function includes information on the number of values that are equal to or greater than a reference value among a plurality of values according to a pass/fail prediction model of M cavities selected from the N cavity groups.
According to claim 11,
Each of the plurality of cavities included in one of the N cavity groups has a higher isotropic than a corresponding plurality of lenses.
According to claim 11,
The plurality of cavities included in one of the N cavity groups are separated from each other after the corresponding plurality of lenses are formed in a state connected to each other before the corresponding plurality of lenses are formed.
A recording medium configured to be accessible to the computing system in order to provide recorded information to the computing system so that the computing system can perform the lens module assembly optimization method of claim 1 or claim 11.

Images