Metal workpiece surface image acquisition method and device based on multiple exposure

Abstract

The invention relates to a metal workpiece surface image acquisition method and device based on multiple exposure. The acquisition method comprises the following steps: determining total exposure, and determining single exposure time and corresponding weight according to the total exposure; acquiring a surface image of the metal workpiece at each exposure time, and adjusting the surface image by using the corresponding weight; and combining a plurality of adjusted images to obtain a metal workpiece surface reflection-removed image. According to the method, a strategy of multiple times of irregular exposure is adopted, multiple times of short-time exposure is used for replacing one time of long-time exposure, a strong single light reflection area is avoided, light reflection is reduced, meanwhile, more image details are reserved, and the final image quality is improved.

Description

Technical Field

[0001] The present invention relates to the field of image acquisition technology, and more specifically to a method and device for acquiring metal workpiece surface images based on multiple exposures, which are used to solve the problem of metal surface reflection and improve image quality. Background Art

[0002] The reflection phenomenon on metal surfaces is a challenge in many application fields, such as high-precision optical measurement, photography, display technology, and manufacturing. Reflections will form bright spots and overexposed areas in the image, making it difficult to identify details, thereby reducing the accurate recognition and measurement of the target object. The reflection phenomenon not only affects the visual quality of the image, but also brings troubles to the subsequent image processing algorithm, causing the loss of important information in the image and affecting the accurate extraction of surface features of the object. In applications such as autonomous driving, industrial inspection, and robot vision, reflections will cause the loss of important information in the image, affecting the accurate extraction of surface features of the object, and thus affecting the performance and reliability of the application.

[0003] At present, the commonly used method to reduce the reflection of metal surfaces includes the use of optical elements such as polarizers. Polarizers reduce the reflected light from the surface by filtering out light with a specific polarization direction. However, this method has significant limitations. The reflection of the metal surface is mainly specular reflection. This reflected light does not change its polarization state during the reflection process, so the polarizer cannot effectively filter out these lights. In addition, the specular reflection of the metal surface is highly directional, and the light is reflected at a specific angle, making the intensity and range of the reflection difficult to predict and control. In practical applications, the variability of the incident angle of light and the complex ambient light conditions further limit the anti-reflection effect of the polarizer. Coupled with the interference of ambient light, the anti-reflection effect of the polarizer becomes even more limited and cannot meet the needs of high-precision imaging.

[0004] Therefore, when collecting images using traditional methods, the loss of image information caused by metal reflection is unavoidable, and the only way is to minimize the impact of metal reflection on imaging. Summary of the invention

[0005] In view of this, the present invention provides a method and device for acquiring metal workpiece surface images based on multiple exposures, aiming to reduce the influence of metal reflection on imaging quality.

[0006] In order to achieve the above object, the present invention adopts the following technical solution:

[0007] On the one hand, the present application provides a method for acquiring a metal workpiece surface image based on multiple exposures, the steps comprising:

[0008] Determine the total exposure, and determine the single exposure time and corresponding weight based on the total exposure;

[0009] Obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight;

[0010] Merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

[0011] Preferably, determining the total exposure includes constructing a total exposure quantification model:

[0012]

[0013] In the formula, E represents the total exposure, I s Indicates the main light source intensity, R s Represents the specular reflection coefficient, cos n (θ) represents the attenuation coefficient of the incident angle, R d represents the diffuse reflection coefficient, α represents the reflection attenuation coefficient, d represents the distance from the main light source to the object, V represents the volume of the irradiated reflective area, I e Indicates the ambient light intensity.

[0014] Preferably, the specular reflection coefficient is obtained by:

[0015]

[0016] In the formula, R f Represents the electromagnetic wave component perpendicular to the incident surface, R p represents the electromagnetic wave component parallel to the incident plane; where,

[0017]

[0018] θ i represents the angle between the incident light and the normal, θ t represents the angle between the refracted light and the normal, n1 represents the refractive index of the incident medium, and n2 represents the refractive index of the projected medium.

[0019] Preferably, the diffuse reflectance is obtained by:

[0020] R d =ρ d cosθ i

[0021] In the formula, ρ d represents the light density, θ i Represents the angle between the incident light ray and the normal.

[0022] Preferably, the single exposure time and the corresponding weight are determined according to the total exposure amount by using the Lagrange multiplier method.

[0023] Preferably, the single exposure time and the corresponding weight are optimized by deep learning, and the steps include:

[0024] The determined single exposure time and the corresponding weight are used as initial parameters, and the parameter feature matrix is ​​extracted, and the image feature matrix of the input image is extracted at the same time; wherein the input image is an image synthesized by the single exposure time and the corresponding weight determined by the Lagrange multiplier method;

[0025] The parameter feature matrix and the image feature matrix are merged and input into the deep learning network to optimize the single exposure time and the corresponding weights.

[0026] Preferably, merging the multiple adjusted images comprises generating a de-reflected image of the metal workpiece surface by a weighted average method:

[0027]

[0028] I' i represents the adjusted image, N represents the number of adjusted images, and i represents the sequence number of the adjusted image.

[0029] In a second aspect, the present application discloses a metal workpiece surface image acquisition device based on multiple exposures, the device executing any of the metal workpiece surface image acquisition methods based on multiple exposures as described above, comprising:

[0030] A total exposure determination unit, used for determining a total exposure;

[0031] An exposure parameter determination unit, used to determine a single exposure time and a corresponding weight according to a total exposure amount;

[0032] An image acquisition unit, used to obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight;

[0033] The image generation unit is used to merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

[0034] Preferably, it also includes:

[0035] An exposure parameter optimization unit, used to extract a parameter feature matrix using the determined single exposure time and the corresponding weight as initial parameters, and simultaneously extract an image feature matrix of an input image; wherein the input image is an image synthesized using the preliminarily determined single exposure time and the corresponding weight;

[0036] And the parameter feature matrix and the image feature matrix are merged and input into the deep learning network to optimize the single exposure time and the corresponding weights.

[0037] The present invention discloses a method and device for acquiring metal workpiece surface images based on multiple exposures. Compared with the prior art, the advantages of the present application are: adopting a strategy of multiple irregular exposures, replacing one long exposure with multiple short exposures, avoiding a strong single reflective area, and retaining more image details while reducing reflections, thereby improving the final image quality.

[0038] Another advantage of the present application is that: an exposure quantification model considering multiple factors is established, and the model can accurately obtain the total exposure of the image;

[0039] Another advantage of the present application is that the exposure time and weight factor are roughly adjusted using the Lagrange multiplier method to ensure the reasonable allocation of each exposure parameter and provide a basis for subsequent fine adjustment;

[0040] Another advantage of this application is that it uses the self-attention mechanism of the Transformer architecture to efficiently process and fuse image features and parameter features to capture complex nonlinear relationships. Through the multi-task learning ability of the neural network model, it can accurately adjust the exposure time and weight factors under different shooting scenes and conditions, thereby significantly reducing the reflection of metal surfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0042] Figure 1 This is a flow chart of a method for acquiring a metal workpiece surface image based on multiple exposures according to the present invention;

[0043] Figure 2 An example diagram of a metal workpiece surface image acquisition method based on multiple exposures of the present invention;

[0044] Figure 3 A flowchart of optimizing exposure parameters based on a deep learning network in the present invention;

[0045] Figure 4 It is a schematic diagram of image synthesis after multiple exposures of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] The embodiment of the present invention discloses a method for acquiring a metal workpiece surface image based on multiple exposures, such as Figure 1 ,include:

[0048] Determine the total exposure, and determine the single exposure time and corresponding weight based on the total exposure;

[0049] Obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight;

[0050] Merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

[0051] In one embodiment, referring to Figure 2 , the execution process of each step is as follows:

[0052] The first step is to build a quantitative model of total exposure and determine the total exposure;

[0053] This embodiment quantifies the total exposure by constructing a total exposure quantification model, and the model form is as follows:

[0054]

[0055] Where E is the total exposure, which refers to the amount of light received by the camera's photosensitive element during the entire exposure process; the total exposure directly affects the brightness and clarity of the captured image;

[0056] I s The main light intensity is the intensity of the main light source shining on the metal surface. The intensity of the main light source determines the brightness of the metal surface reflection.

[0057] R s is the specular reflection coefficient, which indicates the reflective ability of the smooth part of the metal surface. The higher the specular reflection coefficient, the stronger the light reflected by the metal surface;

[0058] cos n (θ) is the attenuation coefficient of the incident angle, θ is the incident angle, n is the attenuation intensity, cos n (θ) represents the effect of the incident angle of the light on the intensity of the reflected light. Generally, the reflected light is stronger when the incident angle is smaller, and the reflected light is weaker when the incident angle is larger;

[0059] R dis the diffuse reflectance, which indicates the reflectivity of the rough part of the metal surface. The higher the diffuse reflectance, the more the light reflected by the metal surface will be scattered in multiple directions;

[0060] α is the reflection attenuation coefficient, which describes the attenuation of light when it is reflected on the metal surface. The larger the attenuation coefficient, the smaller the intensity of the light after reflection;

[0061] d is the distance from the main light source to the object, which is the main factor determining the position of the light source;

[0062] V is the volume of the illuminated reflective area. Considering the reflective characteristics of the metal surface in three-dimensional space, this application calculates the total light intensity of the reflective area by volume integration. This reflects the comprehensive effect of light reflection in different areas on the entire metal surface;

[0063] I e Indicates the intensity of ambient light, the intensity of natural light or other light sources in the shooting environment, usually measured in luminous flux (lumens) or illuminance (lux). Ambient light affects the brightness of the overall image and the reflective effect of the metal surface.

[0064] In the total exposure quantification model disclosed in this embodiment, the volume integral ∫ V Used to calculate the intensity and distribution of light sources at different locations, combined with the light source intensity I s It can more accurately simulate the light reflection in the actual environment. s and diffuse reflectance R d The two main parts of reflection are described separately to ensure that the model can consider both specular reflection and diffuse reflection. The incident angle θ and the reflection attenuation coefficient α are used to simulate the propagation characteristics of light, where cos n (θ) reflects the change of light with the incident angle, e -αd Indicates the energy loss of light during propagation, the ambient light intensity I e The distance d from the light source takes into account the influence of external conditions. Describes the attenuation of light intensity with distance.

[0065] The present invention comprehensively considers the above factors and obtains the amount of light received by the camera photosensitive element during the entire exposure process through comprehensive calculation.

[0066] Furthermore, in this embodiment, the specular reflection coefficient R is calculated s and diffuse reflectance R d The purpose is to more accurately simulate and understand the reflection behavior of light on metal surfaces, so as to effectively reduce the impact of reflection on imaging quality.

[0067] First, the specular reflection coefficient R s Describes the reflection intensity of light on a smooth surface. This reflection follows the law that the angle of reflection is equal to the angle of incidence.s The calculation can quantify the intensity distribution of light in specular reflection and optimize the brightness control during image synthesis.

[0068] Specifically, specular reflection mainly follows the Fresnel reflection law. Assume that the angle between the incident light and the normal is θ i , the angle between the refracted light and the normal is θ t , n1 represents the refractive index of the incident medium, here it represents air, and n2 represents the refractive index of the outgoing medium, here it represents metal. The Fresnel formula can be used to calculate the reflection coefficient R f and R p , where R f Represents the electromagnetic wave component perpendicular to the incident surface, R p Represents the component of the electromagnetic wave parallel to the plane of incidence.

[0069] For R f Quantity, yes

[0070]

[0071] For R p Quantity, yes

[0072]

[0073] The specular reflection coefficient can be expressed as:

[0074]

[0075] Secondly, the diffuse reflectance R d Used to describe the degree of scattering of light on a rough surface. Diffuse reflection causes light to scatter evenly in multiple directions, which helps to simulate ambient lighting conditions more realistically. In this embodiment, diffuse reflection follows Lambert's Cosine Law, which describes the phenomenon of uniform reflection of light on a rough surface. Diffuse reflection coefficient R d Depends on the incident angle θ of the light i , ρ d Represents the light density, and the intensity of diffuse reflection is proportional to the cosine of the normal angle of the incident light, which can be expressed as:

[0076] R d =ρ d cosθ i

[0077] Step 2: Determine the single exposure time and corresponding weight according to the total exposure amount. The present application quantifies the exposure parameters based on the total exposure amount, which can effectively reduce the reflection of the metal surface and improve the imaging quality. In one embodiment, multiple exposures are designed, each with a different exposure time, and the total number of exposures and each exposure time can be determined according to the total exposure amount E.

[0078] In this embodiment, in order to consider the optimization problem of the objective function and the constraint conditions, the Lagrange multiplier method is used to determine the single exposure time and the corresponding weight based on the total exposure amount.

[0079] The Lagrange multiplier method is a mathematical method used to find the extreme value of a function under constrained conditions. By introducing Lagrange multipliers, the constraints can be combined with the objective function and transformed into an unconstrained optimization problem.

[0080] The constraints in the present invention include:

[0081] The total exposure E is fixed;

[0082] Total exposure time T total It should be noted that, since the same set of image sensors is used and the environmental conditions remain unchanged, the total exposure time T total There is a linear relationship with the total exposure E, and their slope and offset are empirical values ​​and adjustable parameters; therefore, in this application, the total exposure time can be determined according to the total exposure;

[0083] Introducing the weight factor w i , reflecting the importance of each exposure;

[0084] The objective function includes:

[0085] Single exposure time T i : The time of each exposure converges to the minimum;

[0086] Total number of exposures N: The total number of exposures, multiple exposures of unequal duration are performed, converging to the minimum number.

[0087] Assume that each exposure time is T i The corresponding reflection intensity is I i , the total reflection intensity is:

[0088]

[0089] in, Both k and p are coefficients, which are determined in advance according to the conditions of image acquisition. Generally, k is set to 1 / N, p is set to N, where N is the number of exposures, and β represents a hyperparameter.

[0090] Furthermore, the Lagrange number multiplication formula is as follows:

[0091]

[0092] In order to find the optimal solution, the partial derivative of each variable in the Lagrangian function L is calculated and set to zero. The specific steps are as follows:

[0093] Calculate partial derivatives of the introduced variables λ, μ, ν and set them to 0, thus obtaining the values ​​of variables λ, μ, ν;

[0094] For each T i and w i To find the partial derivative, the formula is as follows:

[0095]

[0096] get:

[0097]

[0098] By solving the equation system formed by the above expressions, we can get T i and w i The specific value of .

[0099] It should be noted that the single exposure time T i It needs to be within the minimum and maximum exposure time range. If the value obtained exceeds this range, the relevant parameters need to be readjusted and recalculated. The hyperparameter β is set in advance according to the conditions of image acquisition. If the basic conditions do not change, the hyperparameter β will not change this value first.

[0100] In order to further optimize the above technical solution, the single exposure time and corresponding weights are optimized through deep learning.

[0101] Choose to use deep learning to optimize the weight factor w i and exposure time T i , due to its significant advantages in processing complex data and realizing multi-task learning. That is, the deep learning method not only improves the accuracy of parameter setting, but also enhances the adaptability and robustness of the model, ensuring that it can find the optimal exposure combination under different environments and conditions. In addition, the high scalability and versatility of the deep learning model make this architecture widely applicable to other similar optical optimization problems.

[0102] Specifically, the optimization steps include:

[0103] The determined single exposure time and the corresponding weight are used as initial parameters, and a parameter feature matrix is ​​extracted, and an image feature matrix of an input image is extracted at the same time; wherein the input image is an image synthesized by using the preliminarily determined single exposure time and the corresponding weight;

[0104] The parameter feature matrix and the image feature matrix are merged and input into the deep learning network to optimize the single exposure time and the corresponding weights.

[0105] This embodiment performs optimization based on the calculation results of the Lagrangian method, which can promote the rapid convergence of the neural network and effectively reduce the probability of overfitting of the neural network.

[0106] In one embodiment, due to the Transformer architecture, it is able to efficiently process sequence data with its powerful self-attention mechanism and can capture the complex relationship between image features and parameter features; and the architecture is suitable for processing multimodal data (a combination of images and parameters), thereby achieving more accurate multi-task learning. In addition, the encoder layer of the Transformer has the ability to adaptively adjust and can effectively optimize for different input data (such as images and parameters under different shooting conditions); therefore, this embodiment optimizes the exposure parameters based on the neural network of the Transformer architecture.

[0107] The following describes the calculation details in detail through matrix expressions, and analyzes the operations of each main part step by step.

[0108] like Figure 3 ,First, the ResNet50 model is used to extract image features; ,By using the pre-trained ResNet50 model to extract image features, ,it can effectively capture the details and complex features in the image, ,providing a rich information basis for the subsequent parameter ,optimization.

[0109] In this application, the input image is defined as matrix X, and the image feature matrix H is obtained after processing by ResNet50. image ;

[0110]

[0111] B is the batch size, D image is the characteristic dimension of the image.

[0112] Secondly, the parameter vector P (considering that the dimension of the neural network data may change, that is, there may be cropping or 0 padding, so the dimension between the exposure result and P may be different) is processed through the fully connected layer (FC layer) to obtain the parameter feature matrix H param ,The fully connected layer processes various exposure-related parameters, so that these parameters can be fully integrated into the model and improve the prediction ability of the model.

[0113]

[0114] W param is the weight matrix of the parameters, b param is the bias vector of the parameters, D modelIs the dimension of the Transform model.

[0115] Then merge the image features and parameter features, and expand the dimension to fit the Transformer input;

[0116]

[0117] The merged features are then processed through the Transformer encoder to obtain the Transformer output H Transformer ;

[0118]

[0119] Finally, the output of Transformer is converted into exposure time T through a fully connected layer. i and weight factor w i ; The expression is:

[0120] T i =W T H Transformer [:,0,:]+b T

[0121] w i =W W H Transformer [:,0,:]+b w

[0122] The above steps together achieve a comprehensive optimization of exposure parameters, thereby reducing reflections from metal surfaces and improving image quality.

[0123] Step 3: Obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight;

[0124] In this embodiment, the exposure time T i and weight factor w i After that, you can start collecting pictures; T i As an exposure time sequence, it can be transmitted to the image sensor, so that the sensor can collect multiple image sequences with different exposure times at different exposure times. i Both record the details of the same scene under different lighting conditions. By using image sequences with different exposure times, it is possible to capture a wider range of light information in the scene.

[0125] Then, using the resulting weight factor w i , adjust the brightness and transparency of each image. The adjusted image is recorded as I' i , the adjustment formula is:

[0126] I' i=w i I i

[0127] Step 4: Merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

[0128] In one embodiment, Figure 4 , select an image as the base, and then superimpose other images on the image to generate a synthesized image. Preferably, from the multiple exposure image sequence, select an image with the best exposure effect as the base image I base The base image is usually an image with moderate exposure time, good details and brightness. This can provide a stable foundation for subsequent image superposition.

[0129] Finally, all the adjusted images I' i Superimposed on the base image I base The final synthesized high dynamic range (HDR) image I is generated by weighted averaging. final :

[0130]

[0131] Specifically, this means that the brightness and transparency of each image are optimized according to its pre-calculated weight factor. In this way, the light intensity distribution of the composite image under different exposure conditions is more uniform, reducing overexposure or underexposure. This method of synthesizing multiple exposure images can not only effectively reduce the reflection of the metal surface, but also retain more image details, improve the overall imaging quality, and provide support for high-quality imaging.

[0132] Through the above scheme, it can be seen that the present invention adopts a series of complex and precise steps to quantify the degree of reflection on the surface of the metal workpiece, that is, firstly adopting a strategy of multiple irregular exposures to avoid strong reflections caused by a single long exposure, and the image with a short exposure time integrates the light information at different time points, which can effectively retain more image details. This strategy reduces the impact of reflections on the image by capturing the intensity of multiple scattered light.

[0133] Then, the exposure parameters are roughly adjusted using the Lagrange multiplier method to ensure that reasonable initial parameter values ​​are found under various constraints (such as fixed total exposure time and total exposure amount). Through mathematical models and optimization algorithms, the exposure time and weight factor for each exposure can be preliminarily determined, thus providing a basis for subsequent fine-tuning.

[0134] On this basis, the initial parameters are precisely adjusted using a neural network based on the Transformer architecture. The neural network processes a large amount of training data, learns the complex nonlinear relationship between image features and parameter features, and adaptively adjusts each exposure time and weight factor to achieve the best de-reflection effect.

[0135] This multi-level, multi-step optimization method can significantly reduce the reflection on the surface of metal workpieces in different shooting scenarios and environmental conditions, improve the imaging quality, and make the final imaging results clearer and more accurate, thereby providing reliable technical support and high-precision data foundation in applications such as industrial inspection, autonomous driving, and robot vision.

[0136] In an exemplary embodiment, the present invention further provides a metal workpiece surface image acquisition device based on multiple exposures, the device comprising:

[0137] A total exposure determination unit, used for determining a total exposure;

[0138] An exposure parameter determination unit, used to determine a single exposure time and a corresponding weight according to a total exposure amount;

[0139] An image acquisition unit, used to obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight;

[0140] The image generation unit is used to merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

[0141] As a preferred embodiment, the device also includes:

[0142] The exposure parameter optimization unit is used to use the determined single exposure time and the corresponding weight as initial parameters, extract the parameter feature matrix, and simultaneously extract the image feature matrix of the input image; and merge the parameter feature matrix and the image feature matrix and input them into the deep learning network to optimize the single exposure time and the corresponding weight.

[0143] It should be noted that the steps executed by each unit in the image acquisition device of this embodiment are consistent with the aforementioned image acquisition solution, and therefore, they will not be repeated here.

[0144] Compared with the prior art, the present invention replaces a single long exposure with multiple short exposures, which can effectively reduce strong reflections in a single direction while retaining image details. This method ensures that the overall quality and detail of the image are not affected while reducing reflections.

[0145] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0146] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for acquiring metal workpiece surface images based on multiple exposures, characterized in that: Determine the total exposure, and determine the single exposure time and corresponding weight based on the total exposure; Obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight; Merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

2. The image acquisition method according to claim 1, characterized in that: Determine the total exposure, including building a quantitative model for the total exposure: In the formula, E represents the total exposure, I s Indicates the main light source intensity, R s Represents the specular reflection coefficient, cos n (θ) represents the attenuation coefficient of the incident angle, R d represents the diffuse reflection coefficient, α represents the reflection attenuation coefficient, d represents the distance from the main light source to the object, V represents the volume of the irradiated reflective area, I e Indicates the ambient light intensity.

3. The image acquisition method according to claim 2, characterized in that: The specular reflection coefficient is obtained as follows: In the formula, R f Represents the electromagnetic wave component perpendicular to the incident surface, R p represents the electromagnetic wave component parallel to the incident plane; where, θ i represents the angle between the incident light and the normal, θ t represents the angle between the refracted light and the normal, n1 represents the refractive index of the incident medium, and n2 represents the refractive index of the projected medium.

4. The image acquisition method according to claim 2, characterized in that: The diffuse reflectance is obtained as follows: R d =ρ d cosθ i In the formula, ρ d represents the light density, θ i Represents the angle between the incident light ray and the normal.

5. The image acquisition method according to claim 1, characterized in that: The single exposure time and corresponding weight are determined according to the total exposure through the Lagrange multiplier method.

6. The image acquisition method according to claim 1 or 5, characterized in that: The single exposure time and corresponding weights are optimized through deep learning. The steps include: The determined single exposure time and the corresponding weight are used as initial parameters, and a parameter feature matrix is ​​extracted, and an image feature matrix of an input image is extracted at the same time; wherein the input image is an image synthesized by using a single exposure time determined according to a total exposure amount and a corresponding weight; The parameter feature matrix and the image feature matrix are merged and input into the deep learning network to optimize the single exposure time and the corresponding weights.

7. The image acquisition method according to claim 1, characterized in that: Combining multiple adjusted images includes generating a de-reflected image of the metal workpiece surface by weighted averaging: I' i represents the adjusted image, N represents the number of adjusted images, and i represents the sequence number of the adjusted image.

8. A metal workpiece surface image acquisition device based on multiple exposures, characterized in that: The method for acquiring a metal workpiece surface image based on multiple exposures according to any one of claims 1 to 7 comprises: A total exposure determination unit, used for determining a total exposure; An exposure parameter determination unit, used to determine a single exposure time and a corresponding weight according to a total exposure amount; An image acquisition unit, used to obtain the surface image of the metal workpiece at each exposure time and adjust it using the corresponding weight; The image generation unit is used to merge multiple adjusted images to obtain a de-reflected image of the metal workpiece surface.

9. The image acquisition device according to claim 8, characterized in that: Also includes: An exposure parameter optimization unit, used to extract a parameter feature matrix and an image feature matrix of an input image by taking the determined single exposure time and the corresponding weight as initial parameters; And the parameter feature matrix and the image feature matrix are merged and input into the deep learning network to optimize the single exposure time and the corresponding weights.

Images