Map super-resolution reconstruction method based on deep learning
Through the map super-resolution reconstruction method based on deep learning, image optimization is automated, and the problems of low image processing efficiency and limited quality improvement in the existing technology are solved, efficient and detailed game map processing is achieved, and game development efficiency is improved.
Patent Information
- Application Number
- CN202510114737.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-24
- Publication Date
- 2025-05-16
AI Technical Summary
In the prior art, image processing efficiency is low and image quality improvement is limited, especially when processing high-resolution game maps, efficiency problems are prominent, and traditional tools are difficult to balance the sharpness and nature of images.
Using deep learning-based map super-resolution reconstruction method, automated image optimization is performed through deep learning models, including the combination of feature extraction, self-attention mechanism, upsampling module and traditional image processing to generate high-resolution images.
It significantly improves the quality and detailed performance of game maps, reduces the workload of manual image processing, improves game development efficiency, and provides stable and realistic visual effects in a variety of game scenarios.
Smart Images
Figure CN120013763A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of graphics enhancement technology, and in particular to a texture super-resolution reconstruction method based on deep learning. Background Art
[0002] In the existing technology, image processing has the following shortcomings: (1) Low image processing efficiency: Existing image super-resolution technologies often require a lot of computing resources and processing time, especially when processing high-resolution game textures, the efficiency problem is particularly prominent. In addition, these methods may require users to perform complex operations and parameter adjustments during the processing process, which increases the difficulty of use. (2) Limited improvement in image quality: Traditional image enhancement tools have limited effects in improving texture quality, especially in preserving details and eliminating artifacts. These tools often have difficulty balancing the sharpness and naturalness of the image, resulting in the processed image being either too sharp or with lost details.
[0003] Therefore, it is necessary to provide a texture super-resolution reconstruction method based on deep learning to reduce the workload of manual image processing and improve the efficiency of game development. Summary of the invention
[0004] The purpose of the present invention is to provide a texture super-resolution reconstruction method based on deep learning, which reduces the workload of manual image processing and improves game development efficiency through an automated image optimization process.
[0005] In order to solve the problems existing in the prior art, the present invention provides a texture super-resolution reconstruction method based on deep learning, comprising the following steps:
[0006] S1: Input low-resolution game textures into the deep learning model, which extracts features through the head convolution layer and then obtains local and global features of the image through the multi-scale feature extraction module;
[0007] S2: Calculate the similarity between features, use the self-attention mechanism embedded in the deep learning model to understand the image content and perform effective feature fusion;
[0008] At the same time, after feature extraction is completed, the deep learning model gradually restores the spatial size of the image through the upsampling module, while fusing feature information of different scales;
[0009] S3: Through the convolution operation of the output layer of the deep learning model, the high-dimensional features are mapped back to the original number of image channels to generate a high-resolution image;
[0010] S4: Crop the generated high-resolution image to ensure that the output image is consistent with the size of the original input map;
[0011] S5: Preprocessing and traditional image processing of low-resolution game textures;
[0012] S6: Combining the cropped image, preprocessing and traditional image processing, output the image.
[0013] Optionally, in the deep learning-based map super-resolution reconstruction method, the deep learning model is a Self-Attention Knowledge Distillation Network.
[0014] Optionally, in the deep learning-based texture super-resolution reconstruction method, super-resolution is a process of obtaining a high-resolution image through a series of low-resolution images;
[0015] The low resolution is 800×600 or 1024×768; the high resolution is 1920×1080, 2560×1440 or 3840×2160.
[0016] Optionally, in the deep learning-based map super-resolution reconstruction method, the multi-scale feature extraction module includes a Spatial-Temporal Collaborative Backbone.
[0017] Optionally, in the deep learning-based texture super-resolution reconstruction method, the transposed convolution and skip connection used in the upsampling process are used to retain more detail information.
[0018] Optionally, in the deep learning-based texture super-resolution reconstruction method, traditional image processing includes: sharpening, image enhancement, brightness adjustment, area enhancement, saturation adjustment, magnification selection, and magnification style selection.
[0019] Optionally, in the deep learning-based texture super-resolution reconstruction method, the weights of the deep learning model are initialized.
[0020] Compared with the prior art, the present invention has the following advantages:
[0021] (1) The present invention aims to reduce the workload of manual image processing and improve game development efficiency through an automated image optimization process.
[0022] (2) The super-resolution processing method of the present invention significantly improves the quality of game textures. By combining deep learning models with traditional image processing tools, it achieves enhancement of image resolution and details, greatly improving the efficiency of game development.
[0023] (3) The present invention can provide stable and realistic visual effects in a variety of game scenarios, demonstrating its wide applicability and high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 A flow chart of a method provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0025] The specific implementation of the present invention will be described in more detail below in conjunction with the schematic diagram. The advantages and features of the present invention will become clearer based on the following description. It should be noted that the drawings are all in a very simplified form and are not in exact proportions, and are only used to facilitate and clearly assist in explaining the purpose of the embodiments of the present invention.
[0026] Hereinafter, if the method described herein includes a series of steps, the order in which the steps are presented herein is not necessarily the only order in which the steps may be performed, and some of the steps described may be omitted and / or some other steps not described herein may be added to the method.
[0027] In the existing technology, image processing has the following shortcomings: (1) Low image processing efficiency: Existing image super-resolution technologies often require a lot of computing resources and processing time, especially when processing high-resolution game textures, the efficiency problem is particularly prominent. In addition, these methods may require users to perform complex operations and parameter adjustments during the processing process, which increases the difficulty of use. (2) Limited improvement in image quality: Traditional image enhancement tools have limited effects in improving texture quality, especially in preserving details and eliminating artifacts. These tools often have difficulty balancing the sharpness and naturalness of the image, resulting in the processed image being either too sharp or with lost details.
[0028] In order to solve the problems existing in the prior art, the present invention provides a map super-resolution reconstruction method based on deep learning, wherein super-resolution is the process of obtaining a high-resolution image through a series of low-resolution images; the low resolution is a resolution of 800×600 or 1024×768; the high resolution is a resolution of 1920×1080, 2560×1440 or 3840×2160.
[0029] like Figure 1 As shown, the following steps are included:
[0030] S1: Low-resolution game textures are input into the deep learning model, which extracts features through the head convolution layer and then obtains local and global features of the image through a multi-scale feature extraction module (the multi-scale feature extraction module includes the Spatial-Temporal Collaborative Backbone, i.e., STCB). In this process, the image is downsampled to reduce the spatial size while increasing the feature dimension so that features can be extracted at different scales.
[0031] In one embodiment, the deep learning model is Self-Attention Knowledge Distillation Network (SAKDNNet), and the core idea of SAKDNNet is to improve the performance of deep learning models through the self-attention mechanism. Specifically, SAKDN captures long-distance dependencies in input data through the self-attention mechanism. This deep learning model has broad application prospects in natural language processing, image processing and other fields. Self-attention mechanism: SAKDNNet uses a self-attention mechanism to capture long-distance dependencies in input data. The self-attention mechanism allows deep learning models to consider all positions in the sequence at the same time when processing sequence data, so as to better understand contextual information.
[0032] S2: The self-attention mechanism (SAST) embedded in the deep learning model allows the deep learning model to focus on important feature regions when processing images, which is crucial for preserving and enhancing details in super-resolution output.
[0033] Specifically, by calculating the similarity between features, the self-attention mechanism embedded in the deep learning model is used to understand the image content and perform effective feature fusion;
[0034] At the same time, after feature extraction is completed, the deep learning model gradually restores the spatial size of the image through the upsampling module, while fusing feature information of different scales; the transposed convolution and jump connection used in the upsampling process are used to retain more detail information, thereby generating a high-resolution output image.
[0035] S3: Through the convolution operation of the output layer of the deep learning model, the high-dimensional features are mapped back to the original number of image channels to generate a high-resolution image;
[0036] S4: Crop the generated high-resolution image to ensure that the output image is consistent with the size of the original input map;
[0037] S5: Preprocessing and traditional image processing of low-resolution game textures. Traditional image processing includes: sharpening, image enhancement, brightness adjustment, area enhancement, saturation adjustment, magnification selection, and magnification style selection.
[0038] These functions allow users to fine-tune images according to different needs and preferences to ensure that images remain clear and delicate in high-resolution displays. For example, sharpening can enhance edge details in images to make them clearer and more prominent; image enhancement and brightness adjustment can improve the visual effects and detail performance of images; regional enhancement and saturation adjustment can optimize the details and color intensity of different areas in the image; and the selection of magnification and style allows users to choose the appropriate image magnification strategy according to specific application scenarios.
[0039] S6: Combining the cropped image, preprocessing and traditional image processing, output the image.
[0040] Preferably, the weights of the entire deep learning model are initialized by a specific initialization function to facilitate stable training of the deep learning model. During the training process, appropriate loss functions and optimizers are used to minimize the difference between the reconstructed image and the true high-resolution image, ensuring that the deep learning model can learn an effective super-resolution mapping.
[0041] Furthermore, in order to promote the stable training of deep learning models, the weights of the entire deep learning model need to be initialized through a specific initialization method, and the weight matrix can be initialized by sampling from a Gaussian distribution. A Gaussian distribution with a mean of 0 and a standard deviation of 0.01 is given to ensure that each neuron starts learning from a similar initial state.
[0042] The advantages of the initialized deep learning model are:
[0043] (1) Avoiding symmetry problems: If all weights are initialized to the same value (such as zero), all neurons in the network will perform the same calculations symmetrically, resulting in ineffective learning. By sampling from a Gaussian distribution, the weights of each neuron are slightly different, thus avoiding this symmetry problem and promoting each neuron to learn different features.
[0044] (2) Ensure proper signal propagation: Gaussian distribution initialization ensures that the initial value of the weight is not too large or too small, so that the signal can be transmitted smoothly during forward and backward propagation. Especially in deep networks, if the initial value of the weight is too large, it may cause the activation value to be too large, causing gradient explosion; if it is too small, the activation value may become very close to zero, causing the gradient to disappear. Through reasonable initialization, it can ensure that the signal output by the activation function is in the appropriate range, promoting the stable training of deep learning models.
[0045] (3) Accelerated convergence: Using weights sampled from a Gaussian distribution can enable the neural network to converge quickly in the early stages of training. Especially in deep neural networks, reasonable initialization can significantly improve training efficiency.
[0046] During the training process, the EMS loss function is selected as the loss function to calculate the pixel-level error between the generated image and the real high-resolution image. The perceptual loss uses the pre-trained VGG network (such as VGG19) as a feature extractor to calculate the difference between the feature maps of different layers, extract the difference between the generated image and the real image in high-level features, and evaluate the semantic content of the image, not just the pixel-level difference. The adversarial loss helps the generated image to be more realistic. The goal of the adversarial loss is to make the discriminator unable to distinguish between the generated image and the real image, so that the generator can learn to generate high-quality, realistic images. The optimizer selects the Adam optimizer, which combines the momentum and adaptive learning rate methods to automatically adjust the learning rate of each parameter. The learning rate scheduler is used to dynamically adjust the learning rate, so that the deep learning model maintains a stable convergence speed during the training process, and avoids falling into the local optimum too early, helping to overcome the gradient disappearance and gradient explosion problems.
[0047] Through this method, the present invention can significantly improve the visual quality and detail expression of game textures while maintaining computational efficiency.
[0048] In summary, compared with the prior art, the present invention has the following advantages:
[0049] (1) The present invention aims to reduce the workload of manual image processing and improve game development efficiency through an automated image optimization process.
[0050] (2) The super-resolution processing method of the present invention significantly improves the quality of game textures. By combining deep learning models with traditional image processing tools, it achieves enhancement of image resolution and details, greatly improving the efficiency of game development.
[0051] (3) The present invention can provide stable and realistic visual effects in a variety of game scenarios, demonstrating its wide applicability and high efficiency.
[0052] The above is only a preferred embodiment of the present invention and does not limit the present invention in any way. Any technician in the relevant technical field, without departing from the scope of the technical solution of the present invention, makes any form of equivalent replacement or modification to the technical solution and technical content disclosed in the present invention, which does not depart from the content of the technical solution of the present invention and still falls within the protection scope of the present invention.
Claims
1. A texture super-resolution reconstruction method based on deep learning, characterized in that: The following steps are involved: S1: Input low-resolution game textures into the deep learning model, which extracts features through the head convolution layer and then obtains local and global features of the image through the multi-scale feature extraction module; S2: Calculate the similarity between features, use the self-attention mechanism embedded in the deep learning model to understand the image content and perform effective feature fusion; At the same time, after feature extraction is completed, the deep learning model gradually restores the spatial size of the image through the upsampling module, while fusing feature information of different scales; S3: Through the convolution operation of the output layer of the deep learning model, the high-dimensional features are mapped back to the original number of image channels to generate a high-resolution image; S4: Crop the generated high-resolution image to ensure that the output image is consistent with the size of the original input map; S5: Preprocessing and traditional image processing of low-resolution game textures; S6: Combining the cropped image, preprocessing and traditional image processing, output the image.
2. The method for super-resolution reconstruction of textures based on deep learning according to claim 1, characterized in that: The deep learning model is Self-AttentionKnowledge DistillationNetwork.
3. The texture super-resolution reconstruction method based on deep learning according to claim 1, characterized in that: Super-resolution is the process of obtaining a high-resolution image through a series of low-resolution images; The low resolution is 800×600 or 1024×768; the high resolution is 1920×1080, 2560×1440 or 3840×2160.
4. The method for super-resolution reconstruction of textures based on deep learning according to claim 1, characterized in that: The multi-scale feature extraction module includes the Spatial-Temporal Collaborative Backbone.
5. The method for super-resolution reconstruction of textures based on deep learning according to claim 1, characterized in that: The transposed convolution and skip connections used in the upsampling process are used to retain more detail information.
6. The method for super-resolution reconstruction of textures based on deep learning according to claim 1, characterized in that: Traditional image processing includes: sharpening, image enhancement, brightness adjustment, area enhancement, saturation adjustment, magnification selection, and magnification style selection.
7. The method for super-resolution reconstruction of textures based on deep learning according to claim 1, characterized in that: Initialize the weights of the deep learning model.