A method for obtaining and determining MTF value standards using deep learning image detection

Abstract

This invention provides a method for obtaining and determining MTF (Mean Transmission Factor) standards using deep learning image detection, comprising the following steps: taking multiple shots with known MTF values, acquiring video containing the target to be identified for each shot, inputting the video into a target recognition algorithm to output the target failure rate for each shot; obtaining a corresponding failure rate-MTF curve by fitting the target failure rate and MTF using data fitting, considering the negative correlation between failure rate and MTF, and if the interpolation does not satisfy the negative correlation, then using negative correlation fitting to obtain the MTF-failure rate function; performing image recognition on the shot to be tested to obtain the failure rate, and comparing it with the MTF-failure rate function to obtain the MTF of the shot to be tested; and establishing a standardized standard scene with completely fixed target features as a standard scene library, which contains a sufficient number of standard targets, which can be used to verify the obtained conclusions or directly used to identify the MTF of the target shot.

Description

Technical Field

[0001] This invention relates to the field of image processing and image evaluation technology, specifically a method for obtaining and determining MTF value standards using deep learning image detection. Background Technology

[0002] MTF, short for Modulation Transfer Function, is a concept within the realm of physical optics theory and a crucial indicator of optical system quality. Existing as a function of spatial frequency, MTF offers greater analyzability and authority compared to evaluating imaging system quality based solely on a single digital quantity (resolution, sharpness, etc.). Through years of practice, using image MTF to evaluate the imaging quality of optical systems has gained widespread acceptance and application.

[0003] Currently, commonly used MTF calculation methods both domestically and internationally mainly include: the impact input method, the sine input method, the image interpretation method, the pulse method, and the edge method. Among these, the latter two methods are suitable for post-processing calculations, such as calculating the MTF of in-orbit satellite images. In reality, in many applications, the system's MTF is not solely determined by the optical system; backend circuitry also influences it. Since the image is the final output, calculating the MTF from the image can measure the overall imaging quality of the imaging system. The commonly used pulse and edge methods require finding suitable locations in the image—that is, the locations of the pulse and the edge. Sometimes these locations and information may not exist in the image, and sometimes manual intervention is needed to find them. Furthermore, a high-contrast test chart is required to image the image and obtain the Line Spread Function (LSF), which is then used to perform a Fourier transform to obtain the MTF. This measurement method requires precise calibration of the edge, slit, and radiated beam, necessitating complex equipment. All of these factors contribute to the inability to automatically generate an accurate MTF. Summary of the Invention

[0004] The technical problem solved by this invention is to provide a method for obtaining and determining MTF value standards using deep learning image detection, thereby addressing the problems in the background art. To achieve the above objective, this invention provides the following technical solution: a method for obtaining and determining MTF value standards using deep learning image detection, specifically including the following:

[0005] (1) Take multiple shots with known MTF values, and obtain a video containing a sufficient number of targets for each shot; input the video into the algorithm for recognition, and the target failure rate corresponding to each shot can be obtained;

[0006] (2) The target failure rate and MTF are fitted together to obtain a corresponding failure rate-MTF curve. The negative correlation between failure rate and MTF is considered in the data fitting. If the interpolation does not satisfy the negative correlation, the negative correlation fitting is used to obtain the MTF value-failure rate function.

[0007] (3) Perform image recognition on the lens under test to obtain the failure rate, and compare it with the MTF-failure rate function mapping to obtain the MTF of the lens under test;

[0008] (4) Establish a standardized standard scene with completely fixed target features as a standard scene library; the standard scene library contains a sufficient number of standard targets, which can be used to verify the MTF of the obtained lens under test, or directly used to identify the MTF of the target lens.

[0009] Preferably, the data fitting method in (2) includes Lagrange interpolation, where the target failure rate and MTF are used as input values ​​for interpolation fitting to obtain the MTF-failure rate function.

[0010] Preferably, the data fitting method in (2) includes polynomial interpolation and least squares fitting. If Lagrange interpolation cannot obtain an accurate MTF-failure rate function, then polynomial interpolation can be used to interpolate the target failure rate and MTF as input values ​​to obtain a polynomial function. The MTF-failure rate function can be obtained by fitting the polynomial function using the least squares method. The specific method for obtaining the polynomial function by polynomial interpolation is as follows:

[0011] Let set D n D is the set of subscripts of the point (x, y). n ={0, 1, ..., n-1}, construct n polynomials p j (x), j∈D n For any k∈D n They all have p j (x), B k ={i / i≠k,i∈D n}

[0012] Make

[0013]

[0014] Where x is the MTF, y is the target failure rate, and p k (x) is an (n-1)th degree polynomial, and satisfies p k (x m ) = 0 and p k (x k ) = 1. Finally, we can obtain...

[0015]

[0016] The technical effects and advantages of this invention are as follows: This invention does not require observation of sinusoidal grating images. A good MTF estimate can be obtained from a road image containing a sufficient number of recognizable targets, and this is the final MTF, fully including the back-end circuitry. It is estimated that an accurate result can be obtained with approximately 100,000 targets. This invention provides a new standard for judging lens quality, namely, using MTF values ​​as a correlation to derive a standard for judging lens quality based on failure rate; it provides a suggestion for guiding the production of optical lenses in the intelligent driving industry; it provides a simple, intuitive, and highly iterable link between standards in different industries; the algorithm type is not limited, and it can be a recognition algorithm for vehicles, traffic signs, etc., or a recognition algorithm for faces, pupils, etc. One of the core aspects of our method is recognition + establishing a correspondence, thus possessing extremely strong versatility. Attached Figure Description

[0017] Figure 1 This is a flowchart for obtaining the MTF of the lens under test in the embodiment. Detailed Implementation

[0018] To make the technical means, creative features, workflow, and usage methods of this invention readily understandable and effective, the technical solutions in the embodiments of this invention will be clearly and completely described below in conjunction with the embodiments of this invention. Obviously, the described embodiments are merely a part of the embodiments of this invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort, and other conclusions derived from non-creative extensions, are within the scope of protection of this invention.

[0019] Example

[0020] The method for obtaining and determining MTF value standards using deep learning image detection includes the following:

[0021] (1) Take multiple shots with known MTF values, and obtain a video containing a sufficient number of targets for each shot; input the video into the algorithm for recognition, and the target failure rate corresponding to each shot can be obtained;

[0022] (2) Using the target failure rate and MTF as input values, a corresponding failure rate-MTF curve is obtained using Lagrange interpolation. The negative correlation between failure rate and MTF is considered during data fitting. If the interpolation does not satisfy the negative correlation, negative correlation fitting is used to obtain the MTF-failure rate function. If the accuracy of the failure rate-MTF curve is insufficient, a polynomial interpolation method is used to interpolate the target failure rate and MTF as input values ​​to obtain a polynomial function. The MTF-failure rate function is then obtained by fitting the polynomial function using the least squares method. The specific method for obtaining the polynomial function through polynomial interpolation is as follows:

[0023] Let set D n D is the set of subscripts of the point (x, y). n ={0, 1, ..., n-1}, construct n polynomials p j (x), j∈D n For any k∈D n They all have p j (x), B k ={i / i≠k,i∈D n}

[0024] Make

[0025]

[0026] Where x is the MTF, y is the target failure rate, and p k (x) is an (n-1)th degree polynomial, and satisfies p k (x m ) = 0 and p k (x k ) = 1. Finally, we can obtain...

[0027]

[0028] (3) Perform image recognition on the lens under test to obtain the failure rate. Furthermore, this method does not focus on the type of network or the recognition effect for different deep learning neural networks; it only needs to obtain the difference in failure rates to draw a conclusion. For example, using a CNN (Convolutional Neural Network) network and road vehicles as the scene and target, the following definitions are made: 1. Tires and roof are visible; 2. The line connecting the tire contact point and the roof is unobstructed; 3. The vehicle is not being transported or is a virtual vehicle in the image. Vehicles that meet the definitions are "valid targets," and the total number is the number of identifiable targets. After inputting into the network, the number of targets correctly identified by the algorithm is the "number of identified targets." Recognition rate = (number of identified targets / number of valid targets) × 100%. Failure rate η = 1 - recognition rate. Then, compare the obtained failure rate with the MTF-failure rate function mapping to obtain the MTF of the lens under test.

[0029] (4) Establish a standardized standard scene with completely fixed target features as a standard scene library. This library contains a sufficient number of standard targets, which can be used to verify the MTF of the obtained lens under test, or directly used to identify the MTF of the obtained target lens, and the recognition rate during the standard library verification is required to have an error of no more than 0.2% compared with the initial recognition rate.

[0030] Furthermore, this method can be used to systematically evaluate the image quality required for target recognition technology. For example, in the autonomous driving industry, a certain algorithm requires a recognition rate of 99.9%. In this case, the MTF value of the lens is 0.80. That is, when launching the current generation of products, lenses with an MTF ≥ 0.80 should be selected, and this should be used as the product selection standard. Conversely, a lens with an MTF = 0.80 can also be used to test another algorithm. If the recognition rate is lower than 99.9%, this can also be used as a standard for evaluating the algorithm model.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for obtaining and determining MTF value standards using deep learning image detection, characterized in that, Specifically, it includes the following: (1) Take multiple shots with known MTF values, and obtain a video containing a sufficient number of targets for each shot; input the video into the algorithm for recognition, and the target failure rate corresponding to each shot can be obtained; (2) The target failure rate and MTF are fitted together to obtain a corresponding failure rate-MTF curve. The negative correlation between failure rate and MTF is considered in the data fitting. If the interpolation does not satisfy the negative correlation, the negative correlation fitting is used to obtain the MTF value-failure rate function. (3) Perform image recognition on the lens under test to obtain the failure rate, and compare it with the MTF-failure rate function mapping to obtain the MTF of the lens under test; (4) Establish a standardized standard scene with completely fixed target features as a standard scene library. This library contains a sufficient number of standard targets, which can be used to verify the MTF of the obtained lens under test, or directly used to identify the MTF of the target lens. Among them, the data fitting method in (2) includes the Lagrange interpolation method, which uses the target failure rate and MTF as input values ​​to perform interpolation fitting to obtain the MTF value-failure rate function; The data fitting methods in (2) include polynomial interpolation and least squares fitting. If Lagrange interpolation cannot obtain an accurate MTF-failure rate function, then polynomial interpolation can be used to interpolate the target failure rate and MTF as input values ​​to obtain a polynomial function. The MTF-failure rate function can be obtained by fitting the polynomial function using the least squares method. The specific method for obtaining the polynomial function by polynomial interpolation is as follows: Let set D n D is the set of subscripts of the point (x, y). n ={0, 1, ..., n-1}, construct n polynomials p j (x), j∈D n For any k∈D n They all have p j (x), B k ={i / i≠k,i∈D n } makes Where x is the MTF, y is the target failure rate, and p k (x) is an (n-1)th degree polynomial, and satisfies And p k (x k ) = 1, and finally we can get

Images