An image processing technology optimization method in AR interaction scenarios

Abstract

The present invention discloses an image processing technology optimization method in an AR interaction scene, which belongs to the field of image processing technology, including automatically adjusting filtering parameters; preprocessing color images, using a combination of adaptive median filtering and mean filtering, and then using an amplitude-based weight pruning MobileNetV2 model to further remove noise in the color image; graying the color image after noise removal, and limiting the height of the local histogram according to the illumination and details of different areas in the grayscale image; reconstructing the image based on the grayscale image; constructing an FSRCNN model, processing the reconstructed image, and enhancing the resolution of the low-resolution image; reintegrating the processed image resources, and updating them to the metaverse interaction. The present invention adopts the above method, which can significantly improve the image quality, retain image details, and provide users with a richer and more immersive interactive experience.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to an image processing technology optimization method in an AR interaction scenario. Background Art

[0002] In AR interactive scenes, image processing technology plays a vital role to achieve highly realistic user experience and interactive effects. However, with the continuous expansion of AR applications and the continuous improvement of users' requirements for interactive experience, existing image processing methods still have the following problems: the existing technology has limited effect in processing complex noise types, and it is difficult to effectively remove different types of noise, which easily leads to image quality degradation. In order to achieve a certain denoising effect, image details are often lost; in AR scenes, accurate edge information is crucial for the fusion of virtual objects and real scenes. The existing bilateral filtering faces the current situation of difficult parameter adjustment when alleviating the contradiction between edge preservation and smoothing; the existing technology cannot perform adaptive color and detail adjustments according to the content of the image and the characteristics of the scene, resulting in color deviation, blurred details and other problems in the enhanced image; the existing super-resolution reconstruction method lacks effective use of other information in the AR scene, resulting in unnatural transitions between different levels of the reconstructed buildings, loss of realism, and limiting the diversity and richness of AR interactive content. Summary of the invention

[0003] The purpose of the present invention is to provide an image processing technology optimization method in an AR interaction scenario to solve the problems arising in the technical background.

[0004] To achieve the above object, the present invention provides an image processing technology optimization method in an AR interaction scenario, the steps comprising:

[0005] Use deep learning algorithms to train historical image data, predict and automatically adjust filter parameters in AR interaction scenarios;

[0006] The color images acquired in the AR interaction scene are preprocessed, and the salt and pepper noise in the color images is removed by combining adaptive median filtering with mean filtering.

[0007] Build a MobileNetV2 model and combine it with amplitude-based weight pruning to further remove noise from color images;

[0008] The color image after noise removal is grayed out, and the height of the local histogram is limited by using a local adaptive histogram equalization method combined with contrast limitation according to the illumination and details of different areas in the gray image.

[0009] The Frankle-McCann Retinex algorithm with logarithmic transformation preprocessing was used for image reconstruction;

[0010] Using the nonlinear mapping layer, the FSRCNN model is constructed to process the reconstructed image and enhance the resolution of the low-resolution image;

[0011] Reintegrate the processed image resources and update them to the Metaverse interactive scene.

[0012] Preferably, the method of removing salt and pepper noise in a color image by combining adaptive median filtering with mean filtering includes:

[0013] According to the adjusted initial filter window, 5×5 pixels are used to traverse the color image pixels;

[0014] Calculate the median, maximum, and minimum values ​​of the pixel values ​​in the filter window, and choose whether to expand the filter window size according to the pixel value type until the window center value is no longer the maximum or minimum or the window size reaches the preset upper limit.

[0015] Preferably, the constructing of the MobileNetV2 model, combined with amplitude-based weight pruning, to further remove noise in the color image comprises:

[0016] Build a MobileNetV2 model, resize the images after removing salt and pepper noise to the pixel size required by the model, and normalize them;

[0017] Initialize the MobileNetV2 model and modify the model output layer to output an image with the same size and number of channels as the input image;

[0018] Select mean square error as the loss function to calculate the pixel difference between the denoised image output by the model and the true noise-free image;

[0019] The normalized image is input into the MobileNetV2 model for training. The predicted denoised image is obtained through forward propagation. The loss value is calculated according to the mean square error loss function. The Adam optimization algorithm is used to calculate the gradient according to the loss value and update the model.

[0020] After training is completed, analyze the magnitude of each layer's weights, test different thresholds, observe the performance of the pruned model on the validation set, and determine the threshold;

[0021] Perform pruning to reset the weights whose absolute values ​​are less than the threshold to 0, obtain a sparse weight matrix, and reduce the number of model parameters and the amount of calculation.

[0022] Preferably, the method of using a local adaptive histogram equalization method combined with limiting contrast according to the illumination and detail conditions of different regions in the grayscale image, and limiting the height of the local histogram includes:

[0023] Calculate the brightness mean, standard deviation and gradient information of the grayscale image, and determine the illumination boundary through the horizontal and vertical gradients;

[0024] Divide the image blocks according to illumination and detail analysis, calculate the grayscale histogram and cumulative distribution function of each image block, and convert the original grayscale value into the equalized grayscale value;

[0025] Applies contrast-constrained adaptive histogram equalization, clipping the excess to limit the height of the local histogram.

[0026] Preferably, the gray value calculation formula is expressed as:

[0027]

[0028] In the formula, gray is the original gray value, CDF min is the minimum value of the cumulative distribution function, n is the total number of pixels in the small block, L is the total number of gray levels, new_gray is the new gray value, and CDF(gray) is the cumulative distribution function corresponding to each gray value.

[0029] Preferably, the image reconstruction using the Frankle-McCann Retinex algorithm with logarithmic transformation preprocessing includes:

[0030] Perform logarithmic transformation on the intensity value of each pixel in the grayscale histogram;

[0031] Taking the pixel point (x, y) as the center, calculate the weighted sum of it and the surrounding pixels according to the Gaussian center-surround function;

[0032] Calculate the reflection component of the pixel point and perform exponential transformation based on the obtained reflection component;

[0033] The processed grayscale image is reconstructed and the grayscale image is directly output after reconstruction.

[0034] Preferably, the pixel logarithmic transformation formula is expressed as:

[0035] I log (x,y)=log(1+I(x,y));

[0036] Among them, I(x,y) is the pixel value of the original image at the coordinate (x,y), I log (x,y) is the pixel value after logarithmic transformation.

[0037] Preferably, calculating the reflection component of the pixel point and performing exponential transformation according to the obtained reflection component includes:

[0038] The reflection component of the pixel is calculated according to the logarithmic domain formula log(I(x,y))=log(L(x,y))+log(R(x,y)). log (x,y))-R(x,y) performs an exponential transformation on the obtained reflection component to obtain log(R(x,y)), where R(x,y) represents the reflection component of the pixel point (x,y) and L(x,y) represents the illumination component of the pixel point (x,y).

[0039] Preferably, the nonlinear mapping layer is used to construct the FSRCNN model, and the reconstructed image is processed to enhance the resolution of the low-resolution image, including:

[0040] Preprocess the reconstructed image using cropping and scaling operations;

[0041] Constructing an FSERCNN model, wherein the FSERCNN model includes a feature extraction layer, a contraction layer, a nonlinear mapping layer, an expansion layer and a reconstruction layer;

[0042] Use the mean square error loss function to measure the difference between the high-resolution output of the network and the true high-resolution image;

[0043] Select the Adam optimization algorithm to update the network parameters;

[0044] Set the training rounds, input the image data into the model in batches, and forward propagate to obtain the predicted high-resolution images.

[0045] Therefore, the present invention adopts the above-mentioned image processing technology optimization method in an AR interaction scenario, which has the following beneficial effects:

[0046] (1) Adaptive median filtering combined with mean filtering and amplitude-based weight pruning of the MobileNetV2 model are used to ensure good denoising performance while ensuring the lightweight of the model. This helps to remove salt and pepper noise and other noises, thereby improving image quality and preserving image details.

[0047] (2) According to the illumination and detail conditions of different areas, a local adaptive histogram equalization method is used, combined with contrast-limited adaptive histogram equalization to limit the height of the local histogram, further optimize the image enhancement effect, and reduce color deviation and detail blur;

[0048] (3) The Frankle-McCann Retinex algorithm with logarithmic transformation preprocessing is used to adjust and reconstruct components at different scales, perform color correction, and reduce the impact of illumination changes;

[0049] (4) The FSERCNN model can enhance the resolution of low-resolution images, making the transition of reconstructed buildings at different levels more natural and enriching the diversity of AR interactive content;

[0050] (5) Use the FSERCNN model and MobileNetV2 model to extract image semantic features, thereby providing more accurate and rich image content and enhancing user immersion and satisfaction.

[0051] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flowchart of an image processing technology optimization method in an AR interaction scenario according to an embodiment of the present invention;

[0053] Figure 2 It is a flowchart of step S3 of an image processing technology optimization method in an AR interaction scenario according to an embodiment of the present invention;

[0054] Figure 3 It is a flow chart of step S6 of an image processing technology optimization method in an AR interaction scenario according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0056] Reference Figure 1-3 The present invention provides an image processing technology optimization method in an AR interaction scenario, the steps comprising:

[0057] S1. Use deep learning algorithms to train historical image data, predict and automatically adjust filter parameters in AR interaction scenarios. Deep learning algorithms are supervised learning and reinforcement learning algorithms.

[0058] S2. Preprocess the color image obtained in the AR interaction scene, and use a combination of adaptive median filtering and mean filtering to remove salt and pepper noise in the color image.

[0059] Specifically, edge detection, normalization and other operations are performed on the color image obtained in the AR interaction scene; according to the initial filter window determined in step S1, 5×5 pixels are used to traverse the color image pixels; the median, maximum and minimum values ​​of the pixel values ​​in the filter window are calculated, and whether to expand the filter window size is selected according to the pixel value type. If the central pixel value is the maximum or minimum value, the filter window size is expanded, and the median, maximum and minimum values ​​are calculated again. The process is repeated until the window center value is no longer the maximum or minimum or the window size reaches the preset upper limit.

[0060] S3. Build the MobileNetV2 model and combine it with amplitude-based weight pruning to further remove noise from color images. Specifically, it includes:

[0061] S31. Import the deep learning library required for model building, build the MobileNetV2 model, determine the input shape, uniformly adjust the image without salt and pepper noise to 224×224 pixels, normalize it, and map the pixel values ​​to [0,1] to improve the stability of model training.

[0062] It should be noted that the image is uniformly resized to 224×224 pixels to meet the model input requirements, but this process needs to pay attention to maintaining the image ratio and content integrity. If simple stretching or compression may cause image deformation, a scaling method that maintains the aspect ratio can be used, such as first determining the scaling ratio of the shorter side when scaling, and then adjusting the longer side according to the ratio. Normalization is important for mapping pixel values ​​to the [0,1] interval. In model training, the pixel value range of different images may be different. If normalization is not performed, it may cause unstable gradient updates during training. Normalization allows the model to converge faster during training and improves the stability and efficiency of training.

[0063] S32. He initializes the MobileNetV2 model and modifies the model output layer to output an image of the same size and number of channels as the input image.

[0064] Specifically, He initialization initializes the weights by randomly sampling from a normal distribution with a mean of 0 and a variance of (the number of input neurons). This initialization method helps alleviate the gradient vanishing problem that may occur when using the ReLU activation function, allowing the network to converge faster in the early stages of training. Since the model may contain multiple layers with ReLU activation functions, He initialization can ensure that the initial values ​​of the weights of these layers are appropriate.

[0065] S33. Select mean square error as the loss function and calculate the pixel difference between the denoised image output by the model and the true noise-free image.

[0066] The mean square error formula is:

[0067]

[0068] Among them, y i is the pixel value of the noise-free image, is the pixel value of the predicted denoised image, and n is the total pixel value of the image.

[0069] S34. Input the normalized image into the MobileNetV2 model for training, obtain the predicted denoised image through forward propagation, and calculate the loss value according to the mean square error loss function.

[0070] S35. Use the Adam optimization algorithm, set the learning rate to 0.001, β1=0.9 and β2=0.999 and other parameters, calculate the gradient according to the loss value and update the model.

[0071] It should be noted that during the optimization process, we should pay close attention to the training status of the model, such as observing whether the loss value is really decreasing steadily. If we find that the loss value fluctuates abnormally, it may be that the learning rate setting is inappropriate, there is a problem with the gradient calculation, or the model is overfitting. We need to check and adjust the corresponding parameters or improve the model in time. At the same time, since the Adam algorithm has multiple hyperparameters, they affect each other, and we may need to jointly debug these hyperparameters to achieve the best model training effect.

[0072] S36. Set the number of training rounds to 80. During the training process, regularly evaluate the model performance on the validation set to prevent overfitting. After the training is completed, analyze the amplitude of the weights of each layer, and try different thresholds to observe the performance of the pruned model on the validation set and determine the threshold.

[0073] Specifically, the weight amplitude of the convolutional layers close to the input layer is relatively small, because they are mainly responsible for extracting relatively simple, local features. These features are widely distributed in the image and the changes are relatively less drastic, so the corresponding weights do not need to be represented by large values. As the number of network layers increases, the weight amplitude of the part close to the output layer, especially the layer used to fuse high-level semantic information, may be relatively large. And within the same layer, the weight amplitude is not evenly distributed.

[0074] It should be noted that the optimal threshold cannot be determined based on a single indicator, but needs to be considered comprehensively, including noise reduction effect and loss value changes. In addition, the determined threshold is not static, and needs to be adjusted appropriately according to new data, new task requirements, etc. in subsequent practical applications or further optimization.

[0075] The process of determining the threshold is as follows: determine a reasonable threshold test range. Within the determined test range, select different thresholds for testing according to a certain step size. For each selected threshold, perform weight pruning on the model, evaluate the performance of the pruned model on the validation set, and focus on recording the changes in indicators related to the denoising effect and the loss value.

[0076] S37, perform pruning, reset the weights whose absolute values ​​are less than the threshold to 0, obtain a sparse weight matrix, and reduce the number of parameters and calculation amount of the model.

[0077] S4, graying the color image after noise removal, and using a local adaptive histogram equalization method combined with contrast limitation to limit the height of the local histogram according to the illumination and details of different areas in the grayscale image. Specifically including:

[0078] S41, calculating the brightness mean, standard deviation and gradient information of the grayscale image, and determining the illumination boundary through the gradients in the horizontal and vertical directions;

[0079] S42, dividing the image blocks according to the illumination and detail analysis, dividing the area with large illumination changes or rich details into smaller blocks, and dividing the area with uniform illumination and few details into larger blocks. Specifically, for high-resolution images, a relatively large block size (such as 16×16 or 32×32 pixels) can be selected to reduce the amount of calculation while maintaining the representativeness of local features; for low-resolution or detailed images, a smaller block (such as 8×8 pixels) can be selected.

[0080] Calculate the grayscale histogram of each image block and count the frequency of each grayscale level in the small block;

[0081] The grayscale histogram cumulative distribution function of each image block is calculated, and the original grayscale value is converted into the equalized grayscale value through mapping.

[0082] The gray value calculation formula is expressed as:

[0083]

[0084] In the formula, gray is the original gray value, CDF min is the minimum value of the cumulative distribution function, n is the total number of pixels in the small block, L is the total number of gray levels, new_gray is the new gray value, and CDF(gray) is the cumulative distribution function corresponding to each gray value.

[0085] S43, applying contrast-limited adaptive histogram equalization, setting a threshold to limit the height of the local histogram, cutting the excess part and evenly distributing it to other gray levels, recalculating the equalization mapping function, and performing equalization processing on the small block again.

[0086] S5. Use the Frankle-McCann Retinex algorithm with logarithmic transformation preprocessing to reconstruct the image. Specifically include:

[0087] S51, performing logarithmic transformation on the intensity value of each pixel in the grayscale histogram.

[0088] The pixel logarithmic transformation formula is expressed as:

[0089] I log (x,y)=log(1+I(x,y));

[0090] Among them, I(x,y) is the pixel value of the original image at the coordinate (x,y), I log (x,y) is the pixel value after logarithmic transformation.

[0091] S52, taking the pixel point (x, y) as the center, and calculating the weighted sum of the pixel point and surrounding pixels according to the Gaussian center-surround function.

[0092] The Gaussian center-surround function formula is:

[0093]

[0094] Where σ is the standard deviation, which determines the scale of the Gaussian center-surround function, and G(x,y) is the Gaussian center-surround function of the pixel point (x,y).

[0095] It should be noted that σ is a key parameter. A smaller σ makes the distribution of the Gaussian function more concentrated, that is, only the pixels very close to the central pixel have higher weights. The weighted sum obtained in this way focuses more on local details, and the change to the image is relatively small, which may be more suitable for retaining fine features such as the edge of the image. A larger σ will make the distribution of the Gaussian function more dispersed, and more surrounding pixels will have higher weights. In this way, the result after weighted summation is more inclined to comprehensive consideration of a larger area, which can effectively smooth the image, but may blur details such as edges.

[0096] S53, calculating the reflection component of the pixel point, and performing exponential transformation according to the obtained reflection component. Specifically:

[0097] The reflection component of the pixel is calculated according to the logarithmic domain formula log(I(x,y))=log(L(x,y))+log(R(x,y)). log (x,y))-R(x,y) performs an exponential transformation on the obtained reflection component to obtain log(R(x,y)), where R(x,y) represents the reflection component of the pixel point (x,y) and L(x,y) represents the illumination component of the pixel point (x,y).

[0098] S54, reconstructing the processed grayscale image, and directly outputting the reconstructed grayscale image.

[0099] S6. Use the nonlinear mapping layer to build the FSRCNN model, process the reconstructed image, and enhance the resolution of the low-resolution image. Specifically include:

[0100] S61, preprocessing the reconstructed image using operations such as cropping and scaling. First, the retained image area is selected through the pixel coordinate range, data enhancement is performed on the selected image, and images are randomly selected for rotation to obtain additional training data.

[0101] It should be noted that cropping may lose some image information. Therefore, before cropping, you need to carefully consider whether the retained area contains all the key information. Scaling will affect the image quality. When enlarging an image, jagged or blurry images may appear; when reducing an image, detailed information may be lost. Correction needs to be performed when necessary.

[0102] S62. Construct an FSERCNN model. The FSERCNN model includes a feature extraction layer, a contraction layer, a nonlinear mapping layer, an expansion layer and a reconstruction layer.

[0103] Specifically, the feature extraction layer uses 5×5 convolution kernels, the number is 56, and the ReLU activation function is added; the contraction layer shrinks the number of feature maps through the 1×1 convolution kernel; the nonlinear mapping layer constructs 4 3×3 convolution layers, each with 12 convolution kernels, and each layer adds a ReLU activation function; the expansion layer uses a 1×1 convolution kernel to restore the number of feature maps from 12 to 56; the reconstruction layer uses a 9×9 deconvolution kernel with a step size of 4 and a padding of 4 for upsampling, enlarging the image by 4 times.

[0104] S63. Use the mean square error loss function to measure the difference between the high-resolution image output by the network and the true high-resolution image.

[0105] S64, select the Adam optimization algorithm to update the network parameters;

[0106] S65. Set training rounds, input image data into the model in batches, and forward propagate to obtain predicted high-resolution images.

[0107] Specifically, 80 rounds of training were performed, and the model performance was evaluated on the validation set (200 images were selected from the original data as the validation set) every 10 rounds. The learning rate was adjusted according to the validation results, and the training was stopped when the validation loss no longer decreased.

[0108] S7. Reintegrate the processed image resources and update them to the metaverse interaction scene.

[0109] Specifically, for the optimized image resources, first check their format and compatibility to ensure that they meet the requirements of the Metaverse interaction field, then sort out the content relevance according to the scenes, functional areas and application scenarios in the interaction field, classify the relevant images, and then build an integration framework or directory. Then connect to the Unity3D platform, upload or replace image resources in sequence according to the integration framework, and after completion, fully test the performance and compatibility, including checking the image loading speed, display on different terminal devices and related interactive functions. Finally, collect user feedback to continuously optimize image resources and enhance the user's visual experience and satisfaction in Metaverse interaction.

[0110] Therefore, the present invention adopts the above-mentioned image processing technology optimization method in an AR interaction scenario to standardize the image optimization processing process, and can optimize the image fusion effect, improve lighting and details, and ultimately bring high-quality visual presentation and high-quality user experience to image processing in AR interaction scenarios.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for optimizing image processing technology in an AR interaction scenario, characterized in that the steps include: Use deep learning algorithms to train historical image data, predict and automatically adjust filter parameters in AR interaction scenarios; The color images acquired in the AR interaction scene are preprocessed, and the salt and pepper noise in the color images is removed by combining adaptive median filtering with mean filtering. Build a MobileNetV2 model and combine it with amplitude-based weight pruning to further remove noise from color images; The color image after noise removal is grayed out, and the height of the local histogram is limited by using a local adaptive histogram equalization method combined with contrast limitation according to the illumination and details of different areas in the gray image. The Frankle-McCann Retinex algorithm with logarithmic transformation preprocessing was used for image reconstruction; Using the nonlinear mapping layer, the FSRCNN model is constructed to process the reconstructed image and enhance the resolution of the low-resolution image; Reintegrate the processed image resources and update them to the Metaverse interactive scene.

2. The image processing technology optimization method in an AR interaction scenario according to claim 1, characterized in that: The method of combining adaptive median filtering with mean filtering to remove salt and pepper noise in a color image includes: According to the adjusted initial filter window, 5×5 pixels are used to traverse the color image pixels; Calculate the median, maximum, and minimum values ​​of the pixel values ​​in the filter window, and choose whether to expand the filter window size according to the pixel value type until the window center value is no longer the maximum or minimum or the window size reaches the preset upper limit.

3. The image processing technology optimization method in an AR interaction scenario according to claim 2, characterized in that: The construction of the MobileNetV2 model, combined with amplitude-based weight pruning, further removes noise from color images including: Build a MobileNetV2 model, resize the images after removing salt and pepper noise to the pixel size required by the model, and normalize them; Initialize the MobileNetV2 model and modify the model output layer to output an image with the same size and number of channels as the input image; Select mean square error as the loss function to calculate the pixel difference between the denoised image output by the model and the true noise-free image; The normalized image is input into the MobileNetV2 model for training. The predicted denoised image is obtained through forward propagation. The loss value is calculated according to the mean square error loss function. The Adam optimization algorithm is used to calculate the gradient according to the loss value and update the model. After training is completed, analyze the magnitude of each layer's weights, test different thresholds, observe the performance of the pruned model on the validation set, and determine the threshold; Perform pruning to reset the weights whose absolute values ​​are less than the threshold to 0, obtain a sparse weight matrix, and reduce the number of model parameters and the amount of calculation.

4. The image processing technology optimization method in an AR interaction scenario according to claim 3 is characterized in that: According to the illumination and detail conditions of different regions in the grayscale image, a local adaptive histogram equalization method combined with contrast limitation is adopted, and limiting the height of the local histogram includes: Calculate the brightness mean, standard deviation and gradient information of the grayscale image, and determine the illumination boundary through the horizontal and vertical gradients; Divide the image blocks according to illumination and detail analysis, calculate the grayscale histogram and cumulative distribution function of each image block, and convert the original grayscale value into the equalized grayscale value; Applies contrast-constrained adaptive histogram equalization, clipping the excess to limit the height of the local histogram.

5. The image processing technology optimization method in an AR interaction scenario according to claim 4, characterized in that: The gray value calculation formula is expressed as: In the formula, gray is the original gray value, CDF min is the minimum value of the cumulative distribution function, n is the total number of pixels in the small block, L is the total number of gray levels, new_gray is the new gray value, and CDF(gray) is the cumulative distribution function corresponding to each gray value.

6. The method for optimizing image processing technology in an AR interaction scenario according to claim 5, characterized in that: The Frankle-McCann Retinex algorithm using logarithmic transformation preprocessing to perform image reconstruction includes: Perform logarithmic transformation on the intensity value of each pixel in the grayscale histogram; Taking the pixel point (x, y) as the center, calculate the weighted sum of it and the surrounding pixels according to the Gaussian center-surround function; Calculate the reflection component of the pixel point and perform exponential transformation based on the obtained reflection component; The processed grayscale image is reconstructed and the grayscale image is directly output after reconstruction.

7. The method for optimizing image processing technology in an AR interaction scenario according to claim 6, characterized in that: The pixel logarithmic transformation formula is expressed as: I log (x,y)=log(1+I(x,y)); Among them, I(x,y) is the pixel value of the original image at the coordinate (x,y), I log (x,y) is the pixel value after logarithmic transformation.

8. The image processing technology optimization method in an AR interaction scenario according to claim 7, characterized in that: The calculating the reflection component of the pixel point and performing exponential transformation according to the obtained reflection component includes: The reflection component of the pixel is calculated according to the logarithmic domain formula log(I(x,y))=log(L(x,y))+log(R(x,y)). log (x,y))-R(x,y) performs an exponential transformation on the obtained reflection component to obtain log(R(x,y)), where R(x,y) represents the reflection component of the pixel point (x,y) and L(x,y) represents the illumination component of the pixel point (x,y).

9. The method for optimizing image processing technology in an AR interaction scenario according to claim 8, characterized in that: The nonlinear mapping layer is used to construct the FSRCNN model, and the reconstructed image is processed to enhance the resolution of the low-resolution image, including: Preprocess the reconstructed image using cropping and scaling operations; Constructing an FSERCNN model, wherein the FSERCNN model includes a feature extraction layer, a contraction layer, a nonlinear mapping layer, an expansion layer and a reconstruction layer; Use the mean square error loss function to measure the difference between the high-resolution image output by the network and the true high-resolution image; Select the Adam optimization algorithm to update the network parameters; Set the training rounds, input the image data into the model in batches, and forward propagate to obtain the predicted high-resolution images.

Images