Non-uniformity correction method and system based on spatial and frequency domain features
By using an infrared image non-uniformity correction model based on the U-Net structure, combined with spatial and frequency domain feature extraction, the problems of image artifacts and detail loss in traditional methods are solved, achieving high-precision non-uniformity correction results.
Patent Information
- Application Number
- CN202511086366.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-05
- Publication Date
- 2025-12-16
- Estimated Expiration
- 2045-08-05
AI Technical Summary
In existing technologies, traditional infrared image processing methods cannot effectively solve problems such as time-varying detector noise, temperature drift effect, non-uniform pixel response, and radiation distortion, resulting in residual artifacts and loss of detail in the corrected image, making it difficult to apply in scenarios such as security monitoring and industrial inspection.
An infrared image non-uniformity correction model based on the U-Net structure is adopted, which combines spatial and frequency domain feature extraction modules. Through iterative training, training data pairs of pseudo-infrared images and infrared background images are generated to achieve image non-uniformity correction.
It significantly improves the accuracy and robustness of non-uniformity correction, completely eliminates vertical stripe noise, protects image texture and edge information, avoids detail loss caused by over-smoothing, and improves image usability.
Smart Images
Figure CN120580403B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of infrared image processing technology, and more specifically to a non-uniformity correction method and system based on spatial and frequency domain features. Background Technology
[0002] Currently, traditional non-uniformity correction methods based on baffle calibration remain the mainstream technology in infrared image processing. This method uses a baffle image acquired when the thermal infrared camera is powered on as a baseline, and then eliminates fixed-pattern noise through pixel-by-pixel subtraction. Its operation is simple and easy to implement in engineering. However, this type of method has significant drawbacks: First, due to the time-varying characteristics of detector noise, the baseline image and real-time noise cannot be perfectly matched, resulting in residual veil-like artifacts in the corrected image, especially in low-contrast areas, where details are severely lost. Second, the detector's response drift (temperature drift effect) caused by changes in ambient temperature can lead to a shift in baseline features, forcing the system to frequently interrupt operation to reacquire the baffle image, severely limiting its application in continuous operation scenarios such as security monitoring and industrial inspection. More seriously, the non-uniformity of pixel response caused by differences in manufacturing processes of infrared focal plane arrays, radiation distortion caused by uneven spatial distribution of atmospheric refractive index, and center-edge response differences caused by the vignetting effect of the optical system are coupled together, making it difficult for traditional methods to establish accurate global correction models. Therefore, how to solve these problems is a problem that urgently needs to be addressed by those skilled in the art. Summary of the Invention
[0003] In view of this, the present invention provides a non-uniformity correction method and system based on spatial and frequency domain characteristics, which overcomes the above-mentioned defects.
[0004] To achieve the above objectives, the present invention adopts the following technical solution:
[0005] A non-uniformity correction method based on spatial and frequency domain features includes the following steps:
[0006] The calibration steps are as follows:
[0007] Obtain the image to be corrected;
[0008] The image to be corrected is input into a trained infrared image non-uniformity correction model, and the corrected image is output.
[0009] The model training steps are as follows:
[0010] A training dataset is constructed based on a visible light dataset, which includes training data pairs of images to be non-uniformity corrected and pseudo-infrared images.
[0011] An initial infrared image non-uniformity correction model is constructed. The initial infrared image non-uniformity correction model adopts a U-Net structure to construct a non-uniformity correction backbone module. The non-uniformity correction backbone module includes a spatial domain feature extraction module and a frequency domain feature extraction module.
[0012] The initial infrared image non-uniformity correction model is iteratively trained using the training dataset until convergence, thus obtaining the infrared image non-uniformity correction model.
[0013] Furthermore, the steps for constructing the training dataset are as follows:
[0014] Convert the visible light images in the visible light dataset into grayscale images;
[0015] The pixel value distribution of the grayscale image is adjusted to simulate thermal infrared characteristics, forming a pseudo-infrared image;
[0016] An infrared background image is superimposed on the pseudo-infrared image to generate an image to be corrected for non-uniformity, and training data pairs are constructed.
[0017] The training dataset is constructed based on the training data.
[0018] Furthermore, the non-uniformity correction backbone module includes: an encoder, a bottleneck layer, and a decoder connected in sequence; the layers of the encoder are connected through a downsampling module; the layers of the decoder are connected through an upsampling module; the encoder and the decoder are connected by a skip connection; the encoder, the bottleneck layer, and the decoder are all equipped with the spatial domain feature extraction module and the frequency domain feature extraction module.
[0019] Furthermore, the data processing steps of the spatial feature extraction module are as follows:
[0020] The input image is processed sequentially through a layer normalization function, a first convolution, a depthwise convolution, an activation function, a multi-weighted channel attention module, and a second convolution to generate preliminary spatial features;
[0021] The preliminary spatial features are superimposed on the input image to obtain the final spatial features.
[0022] Furthermore, the data processing steps in the multi-weighted channel attention module are as follows:
[0023] Multi-scale features are obtained by performing multi-scale convolution on the input features;
[0024] After global pooling of features at each scale, multi-scale channel features are generated through convolution and the Sigmoid function.
[0025] Multi-scale channel features are concatenated and then probabilistically processed to generate probabilistic features;
[0026] The probability features are convolved to obtain the final channel features;
[0027] The final channel features are multiplied by the input features to obtain multi-weighted channel features.
[0028] Furthermore, the frequency domain feature extraction module extracts frequency domain features by combining real fast Fourier transform and inverse real fast Fourier transform.
[0029] Furthermore, the data processing steps of the frequency domain feature extraction module are as follows:
[0030] The input image is processed sequentially through a third convolution, a first batch of normalization functions, and a first ReLU activation function to obtain a preprocessed feature tensor.
[0031] The preprocessed feature tensor is sequentially processed by real fast Fourier transform, fourth convolution, second batch normalization function, second ReLU activation function and real fast Fourier inverse transform to generate the initial frequency domain feature tensor.
[0032] The initial frequency domain feature tensor and the preprocessed feature tensor are superimposed and then convolved to obtain the final frequency domain feature tensor.
[0033] A non-uniformity correction system based on spatial and frequency domain characteristics includes:
[0034] The training set construction module is used to construct a training dataset based on the visible light dataset, wherein the training dataset includes training data pairs of images to be non-uniformity corrected and pseudo-infrared images;
[0035] The model building module is used to build an initial infrared image non-uniformity correction model. The initial infrared image non-uniformity correction model adopts a U-Net structure to build a non-uniformity correction backbone module. The non-uniformity correction backbone module includes a spatial domain feature extraction module and a frequency domain feature extraction module.
[0036] The model training module is used to iteratively train the initial infrared image non-uniformity correction model using the training dataset until convergence, thereby obtaining the infrared image non-uniformity correction model.
[0037] The image correction module is used to acquire the image to be corrected, input the image to be corrected into a trained infrared image non-uniformity correction model, and output the corrected image.
[0038] As can be seen from the above technical solution, the present invention provides a non-uniformity correction method and system based on spatial and frequency domain characteristics, which has the following advantages compared with the prior art:
[0039] 1. By capturing local pixel-level information through the spatial domain feature extraction module and combining it with the global spectral features extracted by the frequency domain feature extraction module, the synergistic effect of local noise suppression and global structure optimization is achieved, significantly improving the accuracy and robustness of non-uniformity correction.
[0040] 2. For persistent vertical stripe noise in infrared images, frequency domain feature analysis can accurately locate periodic noise components. Combined with spatial adaptive filtering, noise can be completely eliminated, avoiding the residual problems common in traditional methods.
[0041] 3. During the denoising process, the fine extraction of local spatial features and multi-scale fusion strategy effectively protect the texture, edge and weak target information in the image, avoid the loss of details caused by excessive smoothing, and improve the usability of the image. Attached Figure Description
[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0043] Figure 1 This is a schematic diagram of the method flow provided by the present invention;
[0044] Figure 2 A schematic diagram of the structure of the infrared image non-uniformity correction model provided by the present invention;
[0045] Figure 3 This is a schematic diagram of the spatial feature extraction module provided by the present invention;
[0046] Figure 4 This is a schematic diagram of the structure of the multi-weighted channel attention module provided by the present invention;
[0047] Figure 5 This is a schematic diagram of the frequency domain feature extraction module provided by the present invention;
[0048] Figure 6(a) shows the single-point correction result of the first image; Figure 6(b) shows the HINet network correction result of the first image; Figure 6(c) shows the Restormer network correction result of the first image; Figure 6(d) shows the MPRNet network correction result of the first image; Figure 6(e) shows the NAFNet network correction result of the first image; Figure 6(f) shows the correction result of the infrared image non-uniformity correction model provided by the present invention on the first image.
[0049] Figure 7(a) shows the single-point correction result of the second image; Figure 7(b) shows the HINet network correction result of the second image; Figure 7(c) shows the Restormer network correction result of the second image; Figure 7(d) shows the MPRNet network correction result of the second image; Figure 7(e) shows the NAFNet network correction result of the second image; Figure 7(f) shows the correction result of the infrared image non-uniformity correction model provided by the present invention on the second image. Detailed Implementation
[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0051] This invention discloses a non-uniformity correction method based on spatial and frequency domain features, such as... Figure 1 As shown, the specific steps include:
[0052] The calibration steps are as follows:
[0053] Step 1: Obtain the image to be corrected;
[0054] Step 2: Input the image to be corrected into the trained infrared image non-uniformity correction model, and output the corrected image;
[0055] The model training steps are as follows:
[0056] Step 3: Construct a training dataset based on the visible light dataset. The training dataset includes training data pairs of the image to be non-uniformity corrected and the pseudo-infrared image.
[0057] Step 4: Construct an initial infrared image non-uniformity correction model. The initial infrared image non-uniformity correction model uses a U-Net structure to construct a non-uniformity correction backbone module, which includes a spatial domain feature extraction module and a frequency domain feature extraction module.
[0058] Step 5: Iteratively train the initial infrared image non-uniformity correction model using the training dataset until convergence, thus obtaining the infrared image non-uniformity correction model.
[0059] In one embodiment, the steps for constructing the training dataset are as follows:
[0060] Convert visible light images in the visible light dataset to grayscale images;
[0061] Adjusting the pixel value distribution of a grayscale image to simulate thermal infrared characteristics creates a pseudo-infrared image;
[0062] An infrared background image is superimposed on a pseudo-infrared image to generate an image to be corrected for non-uniformity, and training data pairs are constructed.
[0063] A training dataset is constructed based on the training data.
[0064] Furthermore, this embodiment uses existing visible light datasets to simulate and generate a large number of pseudo-infrared images, which are then superimposed with infrared background images to obtain training data pairs of the image to be corrected for non-uniformity and the clean pseudo-infrared image. The specific method is as follows:
[0065] First, the images in the visible light dataset are converted into grayscale images. Then, the pixel values of the grayscale images are further adjusted to make them closer to the real thermal infrared images, resulting in pseudo-infrared images. Next, by superimposing the clean pseudo-infrared image with the infrared background image, the image to be non-uniformly corrected is obtained, thus obtaining a training data pair of the image to be non-uniformly corrected and the clean pseudo-infrared image.
[0066] Furthermore, the network structure diagram of the infrared image non-uniformity correction model is as follows: Figure 2 As shown, it mainly includes a spatial domain feature extraction module, a frequency domain feature extraction module, and a non-uniformity correction backbone module.
[0067] The spatial domain feature extraction module is responsible for extracting information of the image in the pixel domain and completing the non-uniformity correction task based on the local information of the input image; the frequency domain feature extraction module focuses on extracting the frequency domain information of the input image and achieving non-uniformity correction based on the global information of the input image.
[0068] In one embodiment, the data processing steps of the spatial feature extraction module are as follows:
[0069] The input image is processed sequentially through a layer normalization function, a first convolution, a depthwise convolution, an activation function, a multi-weighted channel attention module, and a second convolution to generate preliminary spatial features;
[0070] The preliminary spatial features are superimposed on the input image to obtain the final spatial features.
[0071] Furthermore, the structure of the spatial domain feature extraction module is as follows: Figure 3As shown, specifically: the input feature image tensor (i.e., the input image) is processed sequentially through a layer normalization function, a convolution with a kernel size of 1 (i.e., the first convolution), a depthwise convolution with a kernel size of 3, an activation function, a multi-weight channel attention module, and a convolution with a kernel size of 1 (i.e., the second convolution). The resulting feature image tensor (i.e., the preliminary spatial features) is added to the input feature image tensor as the output of the spatial feature extraction module (i.e., the final spatial features). Among them, the first convolution with a kernel size of 1 doubles the channel dimension of the feature image tensor, the second convolution with a kernel size of 1 halves the channel dimension of the feature image tensor, and the remaining modules do not change the dimension of the feature image tensor.
[0072] In one embodiment, the data processing steps in the multi-weighted channel attention module are as follows:
[0073] Multi-scale features are obtained by performing multi-scale convolution on the input features;
[0074] After global pooling of features at each scale, multi-scale channel features are generated through convolution and the Sigmoid function.
[0075] Multi-scale channel features are concatenated and then probabilistically processed to generate probabilistic features;
[0076] The probability features are convolved to obtain the final channel features;
[0077] The final channel features are multiplied by the input features to obtain multi-weighted channel features.
[0078] Furthermore, the structure diagram of the multi-weighted channel attention module in the spatial domain feature extraction module is as follows: Figure 4 As shown, it first performs convolutions on the input feature image tensor (i.e., the input features) with kernel sizes of 3, 5, 7, and 9 to obtain feature image tensors at four scales. , , , Its dimensions are the same as the input image matrix, both being... Then, the data is processed sequentially through pooling, convolution with a kernel size of 1, activation function (ReLU), convolution with a kernel size of 1, and sigmoid function to obtain multi-scale channel features. , , , All of their dimensions are The multi-scale channel features are combined into a matrix using the Concat function. Its dimensions are Then use the Reshape function to adjust the dimensions to... Then, the Softmax function is applied to the second channel. This function transforms the multi-scale channel features into a probability distribution, where the sum of all probabilities is 1. Finally, the Reshape function is used to reshape the output dimension of the Softmax function back to its original value. The final channel features are obtained through a convolution with a kernel size of 1, and their dimension is... The input feature image tensor is multiplied by the final channel features to obtain the output of the multi-weighted channel attention module (i.e., multi-weighted channel features).
[0079] In one embodiment, the frequency domain feature extraction module extracts frequency domain features by combining real fast Fourier transform and inverse real fast Fourier transform.
[0080] In one embodiment, the data processing steps of the frequency domain feature extraction module are as follows:
[0081] The input image is processed sequentially through the third convolution, the first batch of normalization functions, and the first ReLU activation function to obtain the preprocessed feature tensor.
[0082] The preprocessed feature tensor is processed sequentially through real fast Fourier transform, fourth convolution, second batch normalization function, second ReLU activation function and real fast Fourier transform to generate the initial frequency domain feature tensor;
[0083] The initial frequency domain feature tensor and the preprocessed feature tensor are superimposed and then convolved to obtain the final frequency domain feature tensor.
[0084] Furthermore, the structure of the frequency domain feature extraction module is as follows: Figure 5 As shown, specifically: the input feature image tensor (i.e., the input image) is sequentially processed through a convolution with a kernel size of 1, a batch normalization function, and a ReLU activation function to obtain a preprocessed feature tensor. The preprocessed feature tensor is then sequentially processed through a real fast Fourier transform, a convolution with a kernel size of 1, a batch normalization function, a ReLU activation function, and an inverse real fast Fourier transform to obtain a frequency-domain processed feature image tensor (i.e., the initial frequency-domain feature tensor). The preprocessed feature tensor and the frequency-domain processed feature image tensor are added together, and then convolved with a kernel size of 1 to obtain the output feature image tensor (i.e., the final frequency-domain feature tensor). Specifically, after the real fast Fourier transform, the imaginary and real parts are merged along the channel dimension, thus doubling the dimension of the feature image tensor; before the inverse real fast Fourier transform, the imaginary and real parts are split along the channel dimension, thus halving the dimension of the feature image tensor; the remaining modules do not change the dimension of the feature image tensor.
[0085] In one embodiment, the non-uniformity correction backbone module adopts a U-Net structure with a 5-layer architecture. Each layer of the encoder is connected via a downsampling module, and each layer of the decoder is connected via an upsampling module. A skip connection is added between the encoder and decoder. Spatial domain feature extraction and frequency domain feature extraction modules are used in the encoder, decoder, and bottleneck layer. The channel dimensions of each layer are 32, 64, 128, 256, and 512, respectively; the number of modules in the encoder is 2, 2, 4, and 8; the number of modules in the decoder is 2, 2, 2, and 2; and the number of modules in the bottleneck layer is 12.
[0086] Furthermore, the data processing steps are as follows:
[0087] First, the input feature image tensor is convolved with a kernel size of 1, which changes the dimension of the feature image tensor to 32.
[0088] The feature image tensor is divided into two parts along the channel dimension. These two parts are then processed by the spatial domain feature extraction module and the frequency domain feature extraction module, respectively. The two parts are then merged along the channel dimension and passed through the downsampling module. The downsampling module operates using a convolution with a kernel size of 2 and a stride of 2. The downsampling module can reduce the size of the feature image tensor from... become ,in, Indicates the number of channels. Indicates the height of the image, This indicates the width of the image. Using this module, you can achieve the effect of halving the image size and doubling the number of channels.
[0089] Repeat the above operation three times to obtain the output of the four-layer encoder.
[0090] The bottleneck layer divides the feature image tensor into two parts along the channel. These two parts are then processed by the spatial domain feature extraction module and the frequency domain feature extraction module, respectively, and finally merged along the channel dimension.
[0091] The decoder first passes through an upsampling module, which consists of convolutions with a kernel size of 1 and a PixelShuffle module. The upsampling module reduces the image tensor size from [previous size] using convolutions. become Then, after passing through PixelShuffle, it becomes Then, the output of the corresponding layer encoder is superimposed, and the feature image tensor is divided into two parts in the channel dimension. These two parts are processed by the spatial domain feature extraction module and the frequency domain feature extraction module, respectively. Finally, the two parts are merged in the channel dimension.
[0092] Repeat the operation three times to obtain the output of the four-layer decoder.
[0093] Finally, the feature image tensor is convolved with a kernel size of 1 to reduce the dimension of the feature image tensor to 1, resulting in the final output image.
[0094] The workflow of the encoder, decoder, and spatial and frequency domain feature extraction modules in the bottleneck layer of the non-uniformity correction backbone module is as follows: the feature image tensor is divided into two parts in the channel, and these two parts are respectively processed by the spatial and frequency domain feature extraction modules. Finally, the two parts are merged in the channel dimension to obtain the final output tensor.
[0095] The peak signal-to-noise ratio (PSNR) loss function is used as the loss function in the infrared image non-uniformity correction model. The loss function is defined as follows: ,in, Indicates the number of bits in the image. Indicates a clean infrared image. This represents the infrared image output by the model.
[0096] Figures 6(a)-6(f) and 7(a)-7(f) show the comparison results of various methods in different scenarios. The comparison shows that the method in this embodiment can not only completely remove non-uniform noise, but also preserve the details and edges in the image well, effectively improving the image quality.
[0097] This embodiment also discloses a non-uniformity correction system based on spatial and frequency domain features, including:
[0098] The training set construction module is used to build a training dataset based on the visible light dataset. The training dataset includes training data pairs of images to be non-uniformity corrected and pseudo-infrared images.
[0099] The model building module is used to build an initial infrared image non-uniformity correction model. The initial infrared image non-uniformity correction model adopts the U-Net structure to build the non-uniformity correction backbone module, which includes a spatial domain feature extraction module and a frequency domain feature extraction module.
[0100] The model training module is used to iteratively train the initial infrared image non-uniformity correction model using the training dataset until convergence, thus obtaining the infrared image non-uniformity correction model.
[0101] The image correction module is used to acquire the image to be corrected, input the image to be corrected into the trained infrared image non-uniformity correction model, and output the corrected image.
[0102] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0103] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A non-uniformity correction method based on spatial and frequency domain characteristics, characterized in that, The specific steps include: The calibration steps are as follows: Obtain the image to be corrected; The image to be corrected is input into a trained infrared image non-uniformity correction model, and the corrected image is output. The model training steps are as follows: A training dataset is constructed based on a visible light dataset, which includes training data pairs of images to be non-uniformity corrected and pseudo-infrared images. An initial infrared image non-uniformity correction model is constructed. The initial infrared image non-uniformity correction model adopts a U-Net structure to construct a non-uniformity correction backbone module. The non-uniformity correction backbone module includes a spatial domain feature extraction module and a frequency domain feature extraction module. The workflow of the encoder, decoder, and spatial and frequency domain feature extraction modules in the bottleneck layer of the non-uniformity correction backbone module is as follows: the feature image tensor is divided into two parts in the channel, and these two parts are respectively processed by the spatial and frequency domain feature extraction modules. Finally, the two parts are merged in the channel dimension to obtain the final output tensor. The initial infrared image non-uniformity correction model is iteratively trained using the training dataset until convergence, thus obtaining the infrared image non-uniformity correction model. The data processing steps of the spatial feature extraction module are as follows: The input image is processed sequentially through a layer normalization function, a first convolution, a depthwise convolution, an activation function, a multi-weighted channel attention module, and a second convolution to generate preliminary spatial features; The preliminary spatial features are superimposed on the input image to obtain the final spatial features; The data processing steps in the multi-weighted channel attention module are as follows: Multi-scale features are obtained by performing multi-scale convolution on the input features; After global pooling of features at each scale, multi-scale channel features are generated through convolution and the Sigmoid function. Multi-scale channel features are concatenated and then probabilistically processed to generate probabilistic features; The probability features are convolved to obtain the final channel features; The final channel features are multiplied by the input features to obtain multi-weighted channel features.
2. The non-uniformity correction method based on spatial and frequency domain features according to claim 1, characterized in that, The steps for constructing the training dataset are as follows: Convert the visible light images in the visible light dataset into grayscale images; The pixel value distribution of the grayscale image is adjusted to simulate thermal infrared characteristics, forming a pseudo-infrared image; An infrared background image is superimposed on the pseudo-infrared image to generate an image to be corrected for non-uniformity, and training data pairs are constructed. The training dataset is constructed based on the training data.
3. The non-uniformity correction method based on spatial and frequency domain features according to claim 1, characterized in that, The non-uniformity correction backbone module includes: an encoder, a bottleneck layer, and a decoder connected in sequence; the layers of the encoder are connected through a downsampling module; the layers of the decoder are connected through an upsampling module; and the encoder and the decoder are connected by a skip connection.
4. The non-uniformity correction method based on spatial and frequency domain features according to claim 1, characterized in that, The frequency domain feature extraction module extracts frequency domain features by combining real fast Fourier transform and inverse real fast Fourier transform.
5. The non-uniformity correction method based on spatial and frequency domain features according to claim 4, characterized in that, The data processing steps of the frequency domain feature extraction module are as follows: The input image is processed sequentially through a third convolution, a first batch of normalization functions, and a first ReLU activation function to obtain a preprocessed feature tensor. The preprocessed feature tensor is sequentially processed by real fast Fourier transform, fourth convolution, second batch normalization function, second ReLU activation function and real fast Fourier inverse transform to generate the initial frequency domain feature tensor. The initial frequency domain feature tensor and the preprocessed feature tensor are superimposed and then convolved to obtain the final frequency domain feature tensor.
6. A non-uniformity correction system based on spatial and frequency domain characteristics, characterized in that, include: The training set construction module is used to construct a training dataset based on the visible light dataset, wherein the training dataset includes training data pairs of images to be non-uniformity corrected and pseudo-infrared images; The model building module is used to build an initial infrared image non-uniformity correction model. The initial infrared image non-uniformity correction model adopts a U-Net structure to build a non-uniformity correction backbone module. The non-uniformity correction backbone module includes a spatial domain feature extraction module and a frequency domain feature extraction module. The workflow of the encoder, decoder, and spatial and frequency domain feature extraction modules in the bottleneck layer of the non-uniformity correction backbone module is as follows: the feature image tensor is divided into two parts in the channel, and these two parts are respectively processed by the spatial and frequency domain feature extraction modules. Finally, the two parts are merged in the channel dimension to obtain the final output tensor. The data processing steps of the spatial feature extraction module are as follows: The input image is processed sequentially through a layer normalization function, a first convolution, a depthwise convolution, an activation function, a multi-weighted channel attention module, and a second convolution to generate preliminary spatial features; The preliminary spatial features are superimposed on the input image to obtain the final spatial features; The data processing steps in the multi-weighted channel attention module are as follows: Multi-scale features are obtained by performing multi-scale convolution on the input features; After global pooling of features at each scale, multi-scale channel features are generated through convolution and the Sigmoid function. Multi-scale channel features are concatenated and then probabilistically processed to generate probabilistic features; The probability features are convolved to obtain the final channel features; The final channel features are multiplied by the input features to obtain multi-weighted channel features; The model training module is used to iteratively train the initial infrared image non-uniformity correction model using the training dataset until convergence, thereby obtaining the infrared image non-uniformity correction model. The image correction module is used to acquire the image to be corrected, input the image to be corrected into a trained infrared image non-uniformity correction model, and output the corrected image.
Citation Information
Patent Citations
Low-light image enhancement method based on frequency domain and spatial domain perception
CN118674628A