Residual channel attention network image super-resolution method based on ant colony algorithm collaborative optimization

Abstract

The invention relates to an EDSR (Enhanced Depth Super-Resolution) image super-resolution reconstruction method based on particle swarm optimization, and the method comprises the following steps of: (1) carrying out super-resolution reconstruction on an EDSR (Enhanced Depth Super-Resolution) image; the method comprises the following steps: firstly, inputting a low-resolution image data set as a training sample, defining an optimization space containing the number and stage of residual blocks, convolutional layer parameters, an attention module, an up-sampling mode and the like, initializing particle swarm optimization (PSO) parameters, and dynamically constructing a candidate network by particle position coding; a candidate network is dynamically constructed through particle position coding, and a residual block layer, a convolution layer, an attention module and an up-sampling module are sequentially configured. The candidate network is subjected to limited training, and individual and global optimal positions are updated through fitness function evaluation fusing PSNR and model complexity. And finally, a global optimal structure is selected for complete training for low-resolution image reconstruction, the detail reduction capability and the reasoning efficiency are remarkably improved, and a high-quality image is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of image super-resolution reconstruction, and in particular to an image super-resolution method of a residual channel attention network collaboratively optimized by an ant colony algorithm. Background Art

[0002] Image super-resolution technology aims to recover high-resolution details from low-resolution images and has important applications in fields such as medical imaging and remote sensing monitoring. Early methods primarily relied on interpolation algorithms (such as bicubic interpolation) or models based on sparse coding. However, these methods rely on artificially designed prior assumptions and struggle to recover high-frequency information in complex scenes. With the development of deep learning, convolutional neural networks have significantly improved super-resolution performance. For example, the Residual Channel Attention Network (RCAN) proposed by Zhang et al. achieves leading performance on multiple benchmarks by stacking a deep residual structure and a channel-wise attention mechanism (Zhang et al. (2018), Image super-resolution using very deep residual channel attention networks In Proceedings of ECCV.). However, RCAN's fixed architecture has inherent drawbacks: key parameters such as the number of residual modules and the position of the attention layers must be manually preset, making it difficult for the model to adapt to the complexity requirements of different scenarios. This can easily lead to parameter redundancy in medical image reconstruction, while detail is often lost in satellite image processing.

[0003] In order to reduce the dependence on architecture design, neural network architecture search (NAS) technology was introduced to automatically design network structures. However, the NAS method based on reinforcement learning proposed by Zoph et al. consumes tens of thousands of GPU hours of computing resources, and the search space is optimized for classification tasks and cannot be efficiently migrated to computationally intensive super-resolution tasks. (Zoph et al. (2018), Learning transferable architectures for scalable image recognition In Proceedings of CVPR.) Although swarm intelligence algorithms such as the ant colony optimization proposed by Dorigo et al. perform well in combinatorial optimization problems (Dorigo et al. (2006), Ant colony optimization: Artificial ants as a computational intelligence technique IEEE Computational Intelligence Magazine, 1 (4), 28–39.), The review by et al. pointed out that existing research has not yet solved the problem of its application in super-resolution network module-level architecture search, and current applications are still limited to shallow optimization such as learning rate adjustment. ( et al. (2023), Metaheuristics for deep learning: Recent advances and emerging trends Swarm and Evolutionary Computation, 75, 101-203.) Although techniques such as pruning and quantization can partially reduce model size, the underlying architecture still requires manual design, making it difficult to achieve a globally optimal module combination. Therefore, developing low-computational cost, module-level automated architecture optimization methods has become a key challenge to improving the performance and adaptability of super-resolution models. Summary of the Invention

[0004] In response to the above technical problems, the present invention provides an image super-resolution method of a residual channel attention network collaboratively optimized by an ant colony algorithm, aiming to improve the restoration effect of the image super-resolution task by automatically optimizing the module configuration and module order of the network architecture.

[0005] To achieve the above objectives, the present invention provides an image super-resolution method using a residual channel attention network optimized by an ant colony algorithm, comprising the following steps:

[0006] 1. An image super-resolution method based on an ant colony algorithm to optimize a RCAN network architecture, comprising the following steps:

[0007] S1: Input low-resolution image dataset I LR , as training input samples;

[0008] S2: Define the optimization space of the RCAN network architecture, focusing on the high-level feature extraction part. Specifically, define the following optimization dimensions:

[0009] S21: Number of RG layers N RG : Optimizing the number of RG layers balances the depth and computational complexity of the network, affecting the network's feature extraction capabilities;

[0010] S22: Number of residual blocks (RCAB) N RCAB Optimize the number of RCAB modules to balance model expressiveness and computational efficiency. More RCAB modules improve network feature extraction capabilities, but also increase computational complexity and the risk of overfitting.

[0011] S23: RCAB internal configuration: Optimize the parameters within the RCAB module, including the convolution kernel size K, the number of channels C, the residual connection type, the activation function type, and whether to enable the channel attention mechanism. These factors directly affect the efficiency and quality of feature extraction.

[0012] S3: Initialize the parameters of the ant colony algorithm to provide a behavioral basis for ants, enabling them to have the ability to conduct parallel and global searches in complex structured spaces. The parameters include the number of ants N, the maximum number of iterations T, the pheromone volatility coefficient ρ, the heuristic factor weight α, the heuristic function factor weight β, the pheromone increment constant Q, and the initialization of the pheromone matrix. Among them, i and j represent the current state of the ant and the next network module or configuration selected by the ant respectively;

[0013] S4: Each ant makes a selection decision based on the pheromone concentration and heuristic function to build the RCAN network structure x i , the optimization steps are performed in the following order:

[0014] S41: Stage 1: Select the number of RG layers N in the high-level feature extraction part of the RCAN network RG ;

[0015] S42: Second stage: Determine the number N of RCAB modules used in each RG layer RCAB ;

[0016] S43: The third stage: select the specific configuration of each RCAB module, including the convolution kernel size K, the number of channels C, the residual connection type, the activation function type, and whether to enable the channel attention mechanism.

[0017] The decision-making in each selection stage follows the following probability formula:

[0018]

[0019] Among them, i and j represent the current state of the ant and the next network module or configuration selected by the ant, respectively, k represents all possible path choices of the current node i, τ ij represents the pheromone concentration, η ij is the heuristic factor, α and β are the weights of pheromone and heuristic factor;

[0020] Specifically, the inspiration factor η ij At each stage, set the optimization goal:

[0021] The heuristic factor of S41 is set to the inverse of the number of RG layers: Encourage ants to choose fewer RG layers;

[0022] The heuristic factor of S42 is set to the inverse of the number of RCABs: Encourage ants to select fewer RCAB modules;

[0023] S43's inspiration factors:

[0024] Convolution kernel size: Encourage the selection of smaller convolution kernel sizes;

[0025] Number of channels: Encourage the selection of a smaller number of channels;

[0026] Residual connection type: fixed heuristic factor (such as η res =0.8);

[0027] Activation function type: fixed heuristic factor (such as η act =0.8)

[0028] Channel Attention Mechanism: η att =0.5;

[0029] S5: Train the RCAN structure generated by ants and evaluate it based on image reconstruction performance. The performance evaluation function is:

[0030]

[0031] where x i It is the network architecture. I is the restored super-resolution image and the original high-resolution image, PSNR is the peak signal-to-noise ratio, MAX is the maximum value of the pixel, and MSE is the mean square error between the reconstructed image and the original high-resolution image;

[0032] S6: Update the pheromone concentration according to the performance evaluation results. The update formula is:

[0033]

[0034] Where Δτ ij =Q·f(x i ), Q is the pheromone increment constant, f(x i ) is the performance function of the current hyperparameter combination;

[0035] S7: Determine whether the maximum number of iterations T or the convergence condition is met. If so, proceed to the next step S8; otherwise, return to step S4 to continue iteration;

[0036] S8: Select the optimal path of the ant colony, that is, the optimal network architecture x * , obtained through multiple iterations and pheromone optimization;

[0037] S9: According to the optimal network architecture x *Train the final RCAN network and perform LR Perform super-resolution restoration to generate high-resolution images

[0038]

[0039] In summary, the method of the present invention combines the global optimization capability of the ant colony algorithm with the expressive power of the RCAN network to achieve automated optimization of the module-level network architecture, improve the performance of image super-resolution tasks, reduce the cost of manual design, and has broad practical application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 This is a flowchart of an ant colony algorithm-co-optimized residual channel attention network image super-resolution method of the present invention;

[0041] Figure 2 This is a schematic diagram of the RCAN network structure of the present invention. The present invention mainly applies the ant colony algorithm to the high-level feature extraction part;

[0042] Figure 3 This is a diagram of the internal structure of the RCAB module of the present invention, showing in detail the various components of the RCAB module and its workflow; DETAILED DESCRIPTION

[0043] See below Figures 1 to 3 The present invention provides a detailed description of the image super-resolution method of residual channel attention network with collaborative optimization of ant colony algorithm.

[0044] like Figure 1 As shown, in order to obtain better image super-resolution reconstruction results, a residual channel attention network image super-resolution method optimized by ant colony algorithm is provided, which includes the following steps:

[0045] S1: Input low-resolution image dataset I LR , as training input samples. The dataset can be derived from low-resolution images taken in real life, or generated by downsampling high-resolution images. The image resolution can be set according to specific needs. The low-resolution image size usually used for training is 32x32 to 128x128.

[0046] S2: Define the optimization space of the RCAN network architecture, focusing on the high-level feature extraction part. Specifically, define the following optimization dimensions:

[0047] S21: Number of RG layers N RG : Selected range N RG ∈[3,10], that is, the network contains 3 to 10 RG layers, balancing feature extraction capability and computational complexity;

[0048] S22: Number of residual blocks (RCAB) N RCAB : Selected range N RCAB ∈[5,20], that is, 5 to 20 RCAB modules are selected in each RG layer to balance the network's expressive power and computational complexity;

[0049] S23: RCAB internal configuration: Optimize the following parameters inside the RCAB module:

[0050] Convolution kernel size K: 3x3, 5x5, 7x7;

[0051] Number of channels C: The selection range is C∈[64,256], that is, the number of channels can be adjusted from 64 to 256 to meet the needs of feature extraction;

[0052] Residual connection type: simple residual connection, weighted residual connection, dense residual connection, additive residual connection, etc.

[0053] Activation function type: optional Relu, LeakyRelu, PRelu, etc.

[0054] Whether to enable the channel attention mechanism: 0 or 1 can be selected, 0 means disabled, 1 means enabled. When enabled, the channel attention mechanism can automatically adjust the channel weights in the feature map to enhance the learning ability of important features;

[0055] S3: Initialize the parameters of the ant colony algorithm to provide a behavioral basis for the ants, enabling them to have the ability to search in parallel and globally in complex structured spaces. The parameters include:

[0056] The number of ants N∈[30,50], the maximum number of iterations T∈[100,200], the pheromone volatility coefficient ρ∈[0.8,0.9], the heuristic factor weight α∈[1,3], the heuristic function factor weight β∈[1,3], the pheromone increment constant Q∈[10,50], and the initialization pheromone matrix is a small constant value (such as 0.1);

[0057] S4: Each ant makes a selection decision based on the pheromone concentration and heuristic function to build the RCAN network structure x i , the optimization steps are performed in the following order:

[0058] S41: Stage 1: Select the number of RG layers N in the high-level feature extraction part of the RCAN network RG ;

[0059] S42: Second stage: Determine the number N of RCAB modules used in each RG layer RCAB ;

[0060] S43: The third stage: select the specific configuration of each RCAB module, including the convolution kernel size K, the number of channels C, the residual connection type, the activation function type, and whether to enable the channel attention mechanism.

[0061] The decision-making in each selection stage follows the following probability formula:

[0062]

[0063] Among them, i and j represent the current state of the ant and the next network module or configuration selected by the ant, respectively, k represents all possible path choices of the current node i, τ ij represents the pheromone concentration, η ij is the heuristic factor, α and β are the weights of pheromone and heuristic factor;

[0064] Specifically, the inspiration factor η ij At each stage, set the optimization goal:

[0065] The heuristic factor of S41 is set to the inverse of the number of RG layers: Encourage ants to choose fewer RG layers;

[0066] The heuristic factor of S42 is set to the inverse of the number of RCABs: Encourage ants to select fewer RCAB modules;

[0067] S43's inspiration factors:

[0068] Convolution kernel size: Encourage the selection of smaller convolution kernel sizes;

[0069] Number of channels: Encourage the selection of a smaller number of channels;

[0070] Residual connection type: fixed heuristic factor (such as η res =0.8);

[0071] Activation function type: fixed heuristic factor (such as η act =0.8)

[0072] Channel Attention Mechanism: η att =0.5;

[0073] S5: Train the RCAN structure generated by ants and evaluate it based on image reconstruction performance. The performance evaluation function is:

[0074]

[0075] where x i It is the network architecture. I is the restored super-resolution image and the original high-resolution image, PSNR is the peak signal-to-noise ratio, MAX is the maximum value of the pixel, and MSE is the mean square error between the reconstructed image and the original high-resolution image.

[0076] S6: Update the pheromone concentration according to the performance evaluation results. The update formula is:

[0077]

[0078] Where Δτ ij =Q·f(x i ), Q is the pheromone increment constant, f(x i ) is the performance function of the current hyperparameter combination;

[0079] S7: Determine whether the maximum number of iterations T or the convergence condition is met. If so, proceed to the next step S8; otherwise, return to step S4 to continue iteration;

[0080] S8: Select the optimal path of the ant colony, that is, the optimal network architecture x * , obtained through multiple iterations and pheromone optimization;

[0081] S9: According to the optimal network architecture x * Train the final RCAN network and perform LR Perform super-resolution restoration to generate high-resolution images

[0082]

Claims

1. An image super-resolution method based on residual channel attention network optimized by ant colony algorithm, characterized by: The steps include: S1: Input low-resolution image dataset I LR , as training input samples; S2: Define the optimization space of the RCAN network architecture, focusing on the high-level feature extraction part. Specifically, define the following optimization dimensions: S21: Number of RG layers N RG : Optimizing the number of RG layers balances the depth and computational complexity of the network, affecting the network's feature extraction capabilities; S22: Number of residual blocks (RCAB) N RCAB : Optimize the number of RCAB modules to balance model expressiveness and computational efficiency; more RCAB modules improve network feature extraction capabilities, but also increase computational complexity and possible overfitting risks; S23: RCAB internal configuration: Optimize the parameters within the RCAB module, including the convolution kernel size K, the number of channels C, the residual connection type, the activation function type, and whether to enable the channel attention mechanism. These factors directly affect the efficiency and quality of feature extraction. S3: Initialize the parameters of the ant colony algorithm to provide a behavioral basis for ants, enabling them to have the ability to conduct parallel and global searches in complex structured spaces; the parameters include the number of ants N, the maximum number of iterations T, the pheromone volatility coefficient ρ, the heuristic factor weight α, the heuristic function factor weight β, the pheromone increment constant Q, and the initialization of the pheromone matrix Among them, i and j represent the current state of the ant and the next network module or configuration selected by the ant respectively; S4: Each ant makes a selection decision based on the pheromone concentration and heuristic function to build the RCAN network structure x i , the optimization steps are performed in the following order: S41: Stage 1: Select the number of RG layers N in the high-level feature extraction part of the RCAN network RG ; S42: Second stage: Determine the number N of RCAB modules used in each RG layer RCAB ; S43: The third stage: select the specific configuration of each RCAB module, including the convolution kernel size K, the number of channels C, the residual connection type, the activation function type, and whether to enable the channel attention mechanism. The decision-making in each selection stage follows the following probability formula: Among them, i and j represent the current state of the ant and the next network module or configuration selected by the ant, respectively, k represents all possible path choices of the current node i, τ ij represents the pheromone concentration, η ij is the heuristic factor, α and β are the weights of pheromone and heuristic factor; Specifically, the inspiration factor η ij At each stage, set the optimization goal: The heuristic factor of S41 is set to the inverse of the number of RG layers: Encourage ants to choose fewer RG layers; The heuristic factor of S42 is set to the inverse of the number of RCABs: Encourage ants to select fewer RCAB modules; S43's inspiration factors: Convolution kernel size: Encourage the selection of smaller convolution kernel sizes; Number of channels: Encourage the selection of a smaller number of channels; Residual connection type: fixed heuristic factor (such as η res =0.8); Activation function type: fixed heuristic factor (such as η act =0.8) Channel Attention Mechanism: η att =0.5; S5: Train the RCAN structure generated by ants and evaluate it based on image reconstruction performance. The performance evaluation function is: where x i It is the network architecture. I is the restored super-resolution image and the original high-resolution image, PSNR is the peak signal-to-noise ratio, MAX is the maximum value of the pixel, and MSE is the mean square error between the reconstructed image and the original high-resolution image; S6: Update the pheromone concentration according to the performance evaluation results. The update formula is: Where Δτ ij =Q·f(x i ), Q is the pheromone increment constant, f(x i ) is the performance function of the current hyperparameter combination; S7: Determine whether the maximum number of iterations T or the convergence condition is met. If so, proceed to the next step S8; otherwise, return to step S4 to continue iteration; S8: Select the optimal path of the ant colony, that is, the optimal network architecture x * , obtained through multiple iterations and pheromone optimization; S9: According to the optimal network architecture x * Train the final RCAN network and perform LR Perform super-resolution restoration to generate high-resolution images

Images