An image super-resolution method and system based on lemming optimization and deep learning

Abstract

The present invention discloses an image super-resolution method and system based on lemming optimization and deep learning, and relates to the field of image processing. First, high-resolution and low-resolution image groups are generated based on bicubic interpolation to construct an image super-resolution dataset; then, a hybrid perception image super-resolution model based on KSCeption and HABlock is built; then, the hyperparameter combination of the model is iteratively optimized based on the lemming optimization algorithm, and each set of hyperparameters in the iterative process is used to train and verify the model to obtain the optimal hyperparameter combination and image super-resolution model; finally, the image to be processed is input into the optimal model to obtain a super-resolution image. The present invention effectively fuses global and local features through hybrid perception feature extraction and multi-path feature fusion, and at the same time obtains the model with the best reconstruction effect through the lemming optimization algorithm, thereby improving the quality of the super-resolution image.

Description

Technical Field

[0001] The present invention belongs to the field of image processing technology, and specifically relates to an image super-resolution method and system based on lemming optimization and deep learning. Background Art

[0002] Image resolution is a key metric for measuring image detail and clarity. High-resolution images provide richer spatial and temporal information for motion recognition. In motion recognition scenarios such as sports competitions, security monitoring, intelligent transportation, and human posture analysis, accurately capturing subtle changes in the target's movements and trajectory characteristics often relies on clear and detailed image data. Therefore, introducing image super-resolution technology into the field of motion recognition to restore high-resolution images from low-resolution frames, thereby improving the overall performance of behavior classification, posture estimation, and motion prediction, has important application value and research significance.

[0003] Existing image super-resolution technologies often fail to fully capture and balance local and global features, resulting in artifacts and over-smoothing in the reconstructed images. Furthermore, existing methods cannot effectively handle noise in images, making it difficult to recover subtle details. Furthermore, manually adjusting the model's key hyperparameters is not only labor-intensive, but also difficult to obtain the optimal hyperparameter combination and the best-performing model. Therefore, there is an urgent need to design an image super-resolution method that balances global and local features, effectively handles noise, addresses artifacts, over-smoothing, and loss of detail, and accurately and efficiently selects the optimal model hyperparameters to improve the quality of the reconstructed image. Summary of the Invention

[0004] The technical problem to be solved by the present invention is: in view of the problems that existing image super-resolution methods cannot fully capture and balance local and global features, have difficulty in processing noise, resulting in artifacts, excessive smoothing, and loss of details in the reconstructed image, and the problem that manual parameter adjustment methods cannot accurately search for the optimal hyperparameters of the model, an image super-resolution method and system based on lemming optimization and deep learning are provided to enhance the model's ability to balance local and global features, process noise and restore details, and screen out the optimal hyperparameter combination to maximize the quality of the reconstructed image.

[0005] The present invention adopts the following technical solutions to solve the above technical problems:

[0006] The present invention proposes an image super-resolution method based on lemming optimization and deep learning, comprising the following steps:

[0007] S1. Obtain a high-resolution image and use bicubic interpolation to obtain the corresponding low-resolution image to construct an image super-resolution dataset;

[0008] S2. Build a hybrid perception image super-resolution model based on KSCeption and HABlock;

[0009] S3. Use the Lemmings optimization algorithm to iteratively optimize the hyperparameter combination of the hybrid perception image super-resolution model, use each set of hyperparameters in the iterative process to perform model training and verification, and obtain the optimal hyperparameter combination and corresponding image super-resolution model;

[0010] S4. Input the image to be processed into the optimal model obtained in step S3 to obtain a super-resolution image.

[0011] Furthermore, the hybrid perceptual image super-resolution model in step S2 is specifically composed of the following modules:

[0012] The patch embedding module consists of a convolutional layer and a patch embedding layer. The convolutional layer is used to extract the initial features, and the patch embedding layer is used to split the initial features into patches of smaller sizes.

[0013] The hybrid perceptual feature extraction module consists of several KSCeption units and several HABlock units. The KSCeption unit is used to extract hierarchical features in the image, and the HABlock unit is used to extract long-distance dependent features and fine-grained features in the image.

[0014] The multi-path feature fusion module fuses the features output by different paths in the hybrid perception feature extraction module, further integrates global features and local features using full connection, depthwise convolution, PReLU activation and residual operation, and strengthens the key parts of the fused features. Finally, the fused features are sent to the deconvolution layer to generate a high-resolution output image.

[0015] Furthermore, in the hybrid perception feature extraction module, the KSCeption unit includes two hierarchical feature extraction pathways with the same structure. The expression of its feature extraction pathway is as follows:

[0016] = ,

[0017] = ,

[0018] in, represents the input features of the KSCeption unit, represents the output features of kernel-separated depthwise convolution, represents the output features of the feature extraction pathway; and Respectively and Depthwise convolution of size , and Respectively and dilated convolution of different sizes;

[0019] The output features of the two feature extraction paths are added together to obtain the output features of the KSCeption unit. The calculation formula is as follows:

[0020] ,

[0021] in, represents the output features of the KSCeption unit, Represents the output features of the two feature extraction paths.

[0022] Furthermore, in the hybrid perceptual feature extraction module, based on the sliding window multi-head self-attention SW-MSA, a local feature window block LWFT is constructed. The local feature window block LWFT and the spatial attention block SAB are placed in parallel at the output end of the layer normalization to construct a hybrid attention HABlock unit to extract long-distance dependent features and fine-grained features in the image;

[0023] The calculation formula of HABlock unit is as follows:

[0024] ,

[0025] ,

[0026] ,

[0027] in, represents the input features of HABlock, represents the intermediate features after the first layer normalization, represents mixed features, Represents the output features of HABlock; and Representation layer normalization operation, represents the sliding window multi-head self-attention operation, represents the spatial attention operation, Represents a multilayer perceptron.

[0028] The spatial attention block (SAB) consists of a convolutional layer, a PReLU activation layer, and a spatial attention mechanism. The convolutional layer is used to integrate input features, and the spatial attention mechanism is used to improve the response of key spatial areas. It includes global pooling operations, large-scale convolution operations, PReLU and Sigmoid activation operations, and pixel-by-pixel weighting operations.

[0029] The calculation formula of spatial attention block SAB is as follows:

[0030] ,

[0031] ,

[0032] ,

[0033] ,

[0034] in, represents the input features, represents the intermediate features, represents the spatial attention weight, represents the output features; represents a small-size convolution operation, represents a large-scale convolution operation, Represents the pooling operation in the channel dimension.

[0035] Furthermore, the calculation formula of the multipath feature fusion module is as follows:

[0036] ,

[0037] ,

[0038] ,

[0039] in, represents the multi-path splicing feature, represents the intermediate features, Represents the output features after integration; represents the high-resolution reconstructed image; and represents a full connection operation, represents the depthwise convolution operation, Indicates channel splicing operation, Represents the deconvolution operation.

[0040] Furthermore, step S3 of the present invention specifically includes the following steps:

[0041] S301, set the initialization parameters of the lemming population, including the number of individuals , maximum number of iterations , Problem Dimension , the search space of the lemming search agent group is set to the value range of the hyperparameter combination, which includes the number of KSCeption units Number of HABlock units , the proportional coefficient of the loss function , initial learning rate , training batch size ;

[0042] S302: Initialize the lemming population based on the improved chaotic tent mapping, as follows:

[0043] First, randomly initialize the lemming population within the initialization space [0, 1];

[0044] Then, the population is initialized again using the Tent chaotic map with a perturbation factor, expressed as follows: ,

[0045] in represents the position of the lemming individual after random initialization, Represents the position of the lemming individual after reinitialization, is the perturbation factor constant, Indicates A random number between is the inflection point position control parameter of the Tent mapping, ∈(0,1).

[0046] Finally, the positions of individual lemmings are linearly mapped to the search space;

[0047] S303, using the hyperparameter combination of each individual in the lemming population to train and validate the super-resolution model, calculating the structural similarity (SSIM) between the reconstructed image and the real image during the validation process as the fitness of the individual, and recording the position of the individual with the best fitness;

[0048] S304. Calculate the energy factor E of the current iteration. The energy factor includes a decay coefficient that decreases as the number of iterations increases. At the same time, an adaptive energy coefficient is introduced to improve the lemming optimization algorithm. The coefficient is adaptively adjusted based on the position of the optimal individual in the current iteration and the previous iteration. The energy factor is increased when the iterative optimization effect is poor, and is decreased when the iterative optimization effect is good.

[0049] S305: Based on the current energy factor, simulate the behavior of lemmings to update the position of the current search agent population; set the energy threshold , when the energy is sufficient, that is When the energy is insufficient, the long-distance migration behavior of lemmings when food is scarce, or the behavior of lemmings digging caves, is simulated, so as to conduct global exploration in the search space and avoid falling into the local optimal area; when the energy is insufficient, that is, When simulating the behavior of lemmings foraging in a small area in the cave, or simulating the behavior of lemmings avoiding predators, the fine development of local space is carried out;

[0050] S306: Determine whether the maximum number of iterations has been reached If yes, stop the iteration and output the optimal hyperparameter combination and the corresponding super-resolution model weights. Otherwise, return to step S303 to continue the optimization iteration.

[0051] Furthermore, when the image super-resolution model is trained in step S303, the mean absolute error and the mean square error are combined as the loss function, and the overall loss function is The expression is as follows:

[0052] ,

[0053] in, represents the super-resolution reconstructed image output by the model, represents a real high-resolution image, represents the total number of samples, is the proportional coefficient.

[0054] Furthermore, step S304 calculates the energy factor of the current iteration, specifically:

[0055] Set an adaptive energy coefficient , The value is determined by the best fitness of the current iteration Compared with the best fitness of the previous iteration The degree of optimization is determined by The expression is as follows:

[0056] = ,

[0057] ,

[0058] in, and Represent the minimum and maximum values of the adaptive energy coefficient, is the steepness parameter, is the optimization amplitude threshold;

[0059] The energy factor is calculated as follows:

[0060] ,

[0061] in, Indicates the current iteration number, Indicates the maximum number of iterations.

[0062] In addition, the present invention also proposes an electronic system, comprising: at least one processor; and a memory communicatively connected to the at least one processor; the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method steps described in the present invention.

[0063] The present invention adopts the above technical solution, which has the following technical effects compared with the prior art:

[0064] The present invention proposes an image super-resolution method and system based on lemming optimization and deep learning. A multi-path feature fusion module is used to process the features extracted by the hybrid attention unit HABlock, fully capturing and balancing local and global features, and solving the problems of artifacts and over-smoothing. A lightweight hierarchical feature extraction path based on kernel-separated deep convolution and kernel-separated dilated convolution is constructed, and the hierarchical features are fused through a multi-path feature fusion module, which is beneficial for the model to process noise and restore details. In addition, the population initialization and energy factor of the lemming optimization algorithm are improved to make the initial population distribution more diverse, improve the algorithm's ability to balance exploration and development, find the optimal hyperparameter combination based on the improved lemming optimization algorithm, obtain the image super-resolution model with the best performance, and effectively improve the quality of the reconstructed image. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flow chart of the image super-resolution method proposed in the present invention.

[0066] Figure 2 This is the overall structure diagram of the image super-resolution model proposed in this invention.

[0067] Figure 3 This is the KSCeption unit structure diagram proposed in the present invention.

[0068] Figure 4 This is the HABlock unit structure diagram proposed in the present invention.

[0069] Figure 5 This is the spatial attention block structure diagram proposed in this invention.

[0070] Figure 6 This is the structural diagram of the multi-path feature fusion module proposed in the present invention.

[0071] Figure 7 This is a flow chart of the improved lemming optimization algorithm proposed in the present invention. DETAILED DESCRIPTION

[0072] To make the purpose, technical solutions, and advantages of this application more clear, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work are within the scope of protection of this application:

[0073] like Figure 1 As shown, the present invention proposes an image super-resolution method and system based on lemming optimization and deep learning, which specifically includes the following steps:

[0074] S1. Obtain a high-resolution image and use bicubic interpolation to obtain the corresponding low-resolution image to construct an image super-resolution dataset. The dataset is divided into a training set, a validation set, and a test set.

[0075] S2. Build a hybrid perception image super-resolution model based on KSCeption and HABlock, which consists of a patch embedding module, a hybrid perception feature extraction module, and a multi-path feature fusion module;

[0076] S3. Use the Lemming optimization algorithm to iteratively optimize the model's hyperparameter combination, use each set of hyperparameters in the iterative process to train and verify the model, and obtain the optimal hyperparameter combination and image super-resolution model;

[0077] S4. Input the image to be processed into the optimal model in step S3 to obtain a super-resolution image.

[0078] In one embodiment of the present invention, the training set and validation set in step S1 are taken from the DIV2K dataset, and the test set is taken from the Set5, Set14, BSD100, Urban100, and Manga109 datasets. The training set, validation set, and test set contain 800, 100, and 328 images, respectively.

[0079] In one embodiment of the present invention, step S2 is a hybrid perception image super-resolution model based on KSCeption and HABlock proposed by the present invention. After the lemming optimization algorithm is sought, the number of KSCeption units and HABlock units in the model is set to 4 and 6 respectively. Figure 2 As shown, the hybrid perception image super-resolution model specifically includes:

[0080] Patch Embedding Module: The patch embedding module consists of a convolutional layer and a patch embedding layer. The convolutional layer is used to extract initial features, and the patch embedding layer is used to split the initial features into smaller patches to reduce the computational complexity of the subsequent network.

[0081] Hybrid Perceptual Feature Extraction Module: The hybrid perceptual feature extraction module consists of several KSCeption units and several HABlock units. The KSCeption unit is used to extract hierarchical features in the image, and the HABlock unit is used to extract long-range dependent features and fine-grained features in the image.

[0082] Multi-path feature fusion module: This module fuses the features output by different paths in the hybrid perceptual feature extraction module. It uses full connectivity, depthwise convolution, Prelude (PReLU) activation, and residual operations to further integrate global and local features and strengthen key components of the fused features. Finally, the fused features are fed into a deconvolution layer to generate a high-resolution output image.

[0083] Furthermore, if Figure 3 As shown, the KSCeption unit is specifically:

[0084] A lightweight hierarchical feature extraction pathway is constructed based on kernel separation deep convolution KS-DWConv, kernel separation dilated convolution KS-DilatedConv and PReLU activation function, where the kernel separation method converts a The convolution is decomposed into a serial Convolution and Convolution significantly reduces the number of parameters and computational complexity while maintaining the receptive field and expressive power. Deep convolution is used to extract local hierarchical spatial information, while dilated convolution is used to expand the receptive field, obtain local dependencies, and integrate hierarchical spatial information. PReLU is used to improve the network's nonlinear fitting ability for diverse features. The feature extraction pathway is expressed as follows:

[0085] = ,

[0086] = ,

[0087] in, represents the input features of the KSCeption unit, represents the output features of kernel-separated depthwise convolution, represents the output features of the feature extraction pathway; and Respectively and Depthwise convolution of size , and Respectively and Dilated convolution of different sizes.

[0088] The KSCeption unit connects the two feature extraction paths in parallel and adds the output features of the two feature extraction paths to obtain the KSCeption unit output feature. Feature addition is conducive to the gradient flow during training and can alleviate the gradient vanishing problem. Therefore, the calculation formula of the KSCeption unit is as follows:

[0089] ,

[0090] in, represents the output features of the KSCeption unit, and Represents the output features of the two feature extraction paths.

[0091] Furthermore, if Figure 4 As shown, the HABlock unit is specifically:

[0092] Based on spatial attention convolution, a spatial attention block SAB is constructed to activate more input pixels and improve the global perception ability of the model. Figure 5 As shown in Figure 1, the SAB block consists of a convolutional layer, a PReLU activation layer, and a spatial attention mechanism. The convolutional activation layer is used to integrate input features, and the spatial attention mechanism is used to improve the response of key spatial regions. It includes global pooling operations, large-scale convolution operations, PReLU and Sigmoid activation operations, and pixel-by-pixel weighting operations. The calculation formula of the SAB block is as follows: ,

[0093] ,

[0094] ,

[0095] ,

[0096] in, represents the input features, represents the intermediate features, represents the spatial attention weight, represents the output features; represents a small-size convolution operation, represents a large-scale convolution operation, Represents the pooling operation in the channel dimension.

[0097] Based on the sliding window multi-head self-attention SW-MSA, a local feature window block LWFT is constructed to extract long-distance dependent features and fine-grained features in the image. LWFT and SAB are placed in parallel at the output end of the layer normalization to construct a hybrid attention HABlock unit. The calculation formula of the HABlock unit is as follows:

[0098] ,

[0099] ,

[0100] ,

[0101] in, represents the input features of HABlock, represents the intermediate features after the first layer normalization, Represents mixed features, Represents the output features of HABlock; and Representation layer normalization operation, represents the sliding window multi-head self-attention operation, represents the spatial attention operation, Represents a multilayer perceptron.

[0102] Furthermore, if Figure 6 As shown in Figure 2, the multi-path feature fusion module is specifically as follows:

[0103] Using skip connections, the features of multiple paths in the hybrid perceptual feature extraction module are spliced in the channel dimension. Full connections and depthwise convolution operations are performed to highlight key information. Pre-reduction linear unit (PReLU) and full connections are then used to integrate nonlinear features. Finally, a deconvolution layer is used to output a high-resolution reconstructed image. The multi-path feature fusion module can balance global and local features while fusing hierarchical features. The calculation formula for the multi-path feature fusion module is as follows:

[0104] ,

[0105] ,

[0106] ,

[0107] in, represents the multi-path splicing feature, represents the intermediate features, Represents the output features after integration; represents a high-resolution reconstructed image; and represents a full connection operation, represents the depthwise convolution operation, Indicates channel splicing operation, Represents the deconvolution operation.

[0108] In one embodiment of the present invention, step S3 is the improved lemming optimization algorithm proposed by the present invention, referring to Figure 7 As shown, specifically including:

[0109] S301, set the initialization parameters of the lemming population, including the number of individuals , maximum number of iterations , Problem Dimension , the search space of the lemming search agent group is set to the value range of the hyperparameter combination, which includes the number of KSCeption units Number of HABlock units , the proportional coefficient of the loss function , initial learning rate , training batch size .

[0110] S302. This embodiment proposes a lemming population initialization method based on improved chaotic tent mapping. The lemming population initialization method is specifically as follows:

[0111] Randomly initialize the lemming population within the initialization space [0, 1];

[0112] The population is initialized again using the improved Tent chaotic map. A disturbance factor is introduced on the basis of the traditional Tent map to further increase the diversity of the population. The expression is as follows: ,

[0113] in, represents the position of the lemming individual after random initialization, Represents the position of the lemming individual after reinitialization, is the perturbation factor constant, Indicates A random number between is the inflection point position control parameter of the Tent mapping, ∈(0,1).

[0114] Initialization is completed by linearly mapping the positions of individual lemmings to the search space.

[0115] S303. Use the hyperparameter combination of each individual in the lemming population to train and verify the super-resolution model, calculate the structural similarity (SSIM) between the reconstructed image and the real image during the verification process as the fitness of the individual, and record the position of the individual with the best fitness.

[0116] Furthermore, when training the image super-resolution model, the mean absolute error and mean square error are combined as the loss function, and the overall loss function is The expression is as follows:

[0117] ,

[0118] in, represents the super-resolution reconstructed image output by the model, represents a real high-resolution image, represents the total number of samples, is the proportional coefficient.

[0119] Furthermore, in one embodiment of the present invention, when training the image super-resolution model, the batch size is set to 8, the number of iterations for each training is set to 500, and the optimizer is Adam. and Set to 0.90 and 0.99 respectively.

[0120] S304. Calculate the energy factor E of the current iteration. The energy factor includes a decay coefficient that decreases as the number of iterations increases, which helps to balance early development and late exploration. At the same time, an adaptive energy coefficient is introduced to improve the lemming optimization algorithm. This coefficient can be adaptively adjusted according to the position of the optimal individual in the current iteration and the previous iteration. When the iterative optimization effect is poor, the energy factor is increased to help the algorithm jump out of the local optimal area. When the iterative optimization effect is good, the energy factor is reduced to help the algorithm perform fine local development.

[0121] Furthermore, in this embodiment, the energy factor of the current iteration is calculated as follows:

[0122] Introducing an adaptive energy factor calculation method to improve the energy factor calculation method of the Lemming optimization algorithm. Set an adaptive energy coefficient , The value is determined by the best fitness of the current iteration Compared with the best fitness of the previous iteration The degree of optimization is determined by The expression is as follows:

[0123] = ,

[0124] ,

[0125] in, and Represent the minimum and maximum values of the adaptive energy coefficient, is the steepness parameter, taking a small constant value, is the optimization amplitude threshold.

[0126] The improved energy factor calculation method is as follows:

[0127] ,

[0128] in, Indicates the current iteration number, Indicates the maximum number of iterations.

[0129] S305: Based on the current energy factor, simulate the behavior of lemmings to update the position of the current search agent population and set the energy threshold , when the energy is sufficient, that is When the energy is insufficient, the long-distance migration behavior of lemmings when food is scarce, or the behavior of lemmings digging caves, is simulated, so as to conduct global exploration in the search space and avoid falling into the local optimal area; when the energy is insufficient, that is, When simulating the behavior of lemmings foraging in a small area in the cave, or simulating the behavior of lemmings avoiding predators, the local space is finely developed.

[0130] Furthermore, the simulation of lemming behavior updates the location of the current search agent population as follows:

[0131] when And random probability When , the long-distance migration behavior of lemmings when food is scarce is simulated to update the position of the lemming population. The expression of long-distance migration behavior is:

[0132] ,

[0133] in, and denote the positions of the i-th search agent in the t+1th iteration and the tth iteration, respectively. represents the current optimal solution, represents a search agent randomly selected from the population; -1 or 1, controlling the migration direction; is a random number vector representing Brownian motion; The dimension is (1× ) whose elements are random numbers uniformly distributed in the interval [-1, 1];

[0134] when And random probability When , the lemmings digging behavior is simulated to update the position of the lemming population. The expression of digging behavior is:

[0135] ,

[0136] Where L is a random number related to the current iteration number t, which is used to characterize the interaction between lemmings when digging a burrow. The expression of L is:

[0137] ,

[0138] when And random probability When , the behavior of lemmings foraging in a small area in the cave is simulated, and the position of the lemming population is updated using the spiral wrapping mechanism. The expression of the foraging behavior is:

[0139] ,

[0140] in represents the spiral shape of random search during foraging and is calculated as follows:

[0141] ,

[0142] when And random probability When , the position of the lemming population is updated by simulating the behavior of lemmings avoiding predators. The expression of the behavior of avoiding predators is:

[0143] ,

[0144] ,

[0145] in, is the Levy flight function, which is used to simulate the deceptive actions of lemmings during escape. The expression of the Levy flight function is as follows:

[0146] ,

[0147] ,

[0148] in, 、 is a random value between 0 and 1, is a constant equal to 1.5.

[0149] S306: Determine whether the maximum number of iterations has been reached If so, stop the iteration and output the optimal hyperparameter combination and the corresponding super-resolution model weights. Otherwise, return to continue the optimization iteration.

[0150] Comparing the performance of the present invention with other methods on five test data sets shows the superiority of the present invention. The comparison results are shown in Table 1, where the larger the PSNR and SSIM indicators are, the higher the quality of the reconstructed image.

[0151] Table 1: Comparison of test results of the present invention and other methods on 5 test data sets

[0152]

[0153] Example 2: This embodiment proposes an electronic system, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method steps described in the present invention.

[0154] It should be noted that the processing flow of Example 2 corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not fully described in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0155] The program code for implementing the methods of the present application can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations specified in the flow charts and / or block diagrams are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0156] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0157] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that the specific features, structures, materials, or characteristics described in conjunction with that embodiment or example are included in at least one embodiment or example of the present invention. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0158] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention, and such changes and modifications fall within the scope of the invention as claimed.

Claims

1. An image super-resolution method based on lemming optimization and deep learning, characterized in that: The steps include: S1. Obtain a high-resolution image and use bicubic interpolation to obtain the corresponding low-resolution image to construct an image super-resolution dataset; S2. Build a hybrid perception image super-resolution model based on KSCeption and HABlock; S3. Use the Lemmings optimization algorithm to iteratively optimize the hyperparameter combination of the hybrid perception image super-resolution model, use each set of hyperparameters in the iterative process to perform model training and verification, and obtain the optimal hyperparameter combination and corresponding image super-resolution model; S4, inputting the image to be processed into the optimal model obtained in step S3 to obtain a super-resolution image; Wherein, step S3 specifically includes the following steps: S301, set the initialization parameters of the lemming population, including the number of individuals , maximum number of iterations , Problem Dimension , the search space of the lemming search agent group is set to the value range of the hyperparameter combination, which includes the number of KSCeption units Number of HABlock units , the proportional coefficient of the loss function , initial learning rate , training batch size ; S302: Initialize the lemming population based on the improved chaotic tent mapping, as follows: First, randomly initialize the lemming population within the initialization space [0, 1]; Then, the population is initialized again using the Tent chaotic map with a perturbation factor, expressed as follows: , in represents the position of the lemming individual after random initialization, Represents the position of the lemming individual after reinitialization, is the perturbation factor constant, Indicates A random number between is the inflection point position control parameter of the Tent mapping, ∈(0,1); Finally, the positions of individual lemmings are linearly mapped to the search space; S303, using the hyperparameter combination of each individual in the lemming population to train and validate the super-resolution model, calculating the structural similarity (SSIM) between the reconstructed image and the real image during the validation process as the fitness of the individual, and recording the position of the individual with the best fitness; S304. Calculate the energy factor E of the current iteration. The energy factor includes a decay coefficient that decreases as the number of iterations increases. At the same time, an adaptive energy coefficient is introduced to improve the lemming optimization algorithm. The coefficient is adaptively adjusted based on the position of the optimal individual in the current iteration and the previous iteration. The energy factor is increased when the iterative optimization effect is poor, and is decreased when the iterative optimization effect is good. S305: Based on the current energy factor, simulate the behavior of lemmings to update the position of the current search agent population; set the energy threshold , when the energy is sufficient, that is When the energy is insufficient, the long-distance migration behavior of lemmings when food is scarce, or the behavior of lemmings digging caves, is simulated, so as to conduct global exploration in the search space and avoid falling into the local optimal area; when the energy is insufficient, that is, When simulating the behavior of lemmings foraging in a small area in the cave, or simulating the behavior of lemmings avoiding predators, the fine development of local space is carried out; S306: Determine whether the maximum number of iterations has been reached If yes, stop the iteration and output the optimal hyperparameter combination and the corresponding super-resolution model weights. Otherwise, return to step S303 to continue the optimization iteration.

2. The method according to claim 1, wherein The hybrid perceptual image super-resolution model in step S2 is specifically composed of the following modules: The patch embedding module consists of a convolutional layer and a patch embedding layer. The convolutional layer is used to extract the initial features, and the patch embedding layer is used to split the initial features into patches of smaller sizes. The hybrid perceptual feature extraction module consists of several KSCeption units and several HABlock units. The KSCeption unit is used to extract hierarchical features in the image, and the HABlock unit is used to extract long-distance dependent features and fine-grained features in the image. The multi-path feature fusion module fuses the features output by different paths in the hybrid perception feature extraction module, further integrates global features and local features using full connection, depthwise convolution, PReLU activation and residual operation, and strengthens the key parts of the fused features. Finally, the fused features are sent to the deconvolution layer to generate a high-resolution output image.

3. The method according to claim 2, wherein The KSCeption unit includes two hierarchical feature extraction pathways with the same structure. The feature extraction pathway consists of kernel-separated depthwise convolution, kernel-separated dilated convolution, and PReLU activation function, which are expressed as follows: = ， = ， in, represents the input features of the KSCeption unit, represents the output features of kernel-separated depthwise convolution, represents the output features of the feature extraction pathway; and Respectively and Depthwise convolution of size , and Respectively and dilated convolution of different sizes; The output features of the two feature extraction paths are added together to obtain the output features of the KSCeption unit. The calculation formula is as follows: ， in, represents the output features of the KSCeption unit, and Represents the output features of the two feature extraction paths.

4. The method according to claim 2, wherein Based on the sliding window multi-head self-attention SW-MSA, a local feature window block LWFT is constructed. The local feature window block LWFT and the spatial attention block SAB are placed in parallel at the output end of the layer normalization to construct a hybrid attention HABlock unit to extract long-distance dependent features and fine-grained features in the image; The calculation formula of HABlock unit is as follows: ， ， ， in, represents the input features of HABlock, represents the intermediate features after the first layer normalization, represents mixed features, Represents the output features of HABlock; and Representation layer normalization operation, represents the sliding window multi-head self-attention operation, represents the spatial attention operation, Represents a multilayer perceptron.

5. The method according to claim 4, wherein The spatial attention block (SAB) consists of a convolutional layer, a PReLU activation layer, and a spatial attention mechanism. The convolutional layer is used to integrate input features, and the spatial attention mechanism is used to improve the response of key spatial regions. It includes global pooling operations, large-scale convolution operations, PReLU and Sigmoid activation operations, and pixel-by-pixel weighting operations. The calculation formula of spatial attention block SAB is as follows: ， ， ， ， in, represents the input features, represents the intermediate features, represents the spatial attention weight, represents the output features, represents a small-size convolution operation, represents a large-scale convolution operation, Represents the pooling operation in the channel dimension.

6. The method according to claim 2, wherein The calculation formula of the multipath feature fusion module is as follows: ， ， ， in, represents the multi-path splicing feature, represents the intermediate features, Represents the output features after integration; represents the high-resolution reconstructed image; and represents a full connection operation, represents the depthwise convolution operation, Indicates channel splicing operation, Represents the deconvolution operation.

7. The method according to claim 1, wherein When training the image super-resolution model, the mean absolute error and mean square error are combined as the loss function, and the overall loss function is The expression is as follows: ， in, represents the super-resolution reconstructed image output by the model, represents the real high-resolution image, n represents the total number of samples, is the proportional coefficient.

8. The method according to claim 1, wherein: Step S304 calculates the energy factor of the current iteration, specifically: setting an adaptive energy coefficient , The value is determined by the best fitness of the current iteration Compared with the best fitness of the previous iteration The degree of optimization is determined by The expression is as follows: = ， ， in, and Represent the minimum and maximum values of the adaptive energy coefficient, is the steepness parameter, δ is the optimization amplitude threshold; The energy factor is calculated as follows: ， Where t represents the current iteration number, Indicates the maximum number of iterations.

9. An electronic system comprising: at least one processor; and a memory communicatively connected to the at least one processor; characterized in that the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can perform the method steps described in any one of claims 1-8.

Images