Deep learning non-uniformity correction method combined with reference radiation source

Abstract

The invention provides a deep learning non-uniformity correction method combined with a reference radiation source. The deep learning non-uniformity correction method comprises the following steps: step 1, making a data set; step 2, training a network model by using a supervised mode according to the manufactured data set through the network model, and determining network parameters; step 3, using the trained network model to perform non-uniform correction on the to-be-corrected picture to obtain a corrected image; according to the deep learning non-uniformity correction method in combination with the reference radiation source, deep mining is performed on image structure features, image detail features and spatial level features of different scales of a black body image by using a deep network, all effective image information except non-uniformity noise is reserved, and high-quality non-uniformity correction of the black body image is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared image processing, and in particular relates to a deep learning non-uniform correction method combined with a reference radiation source. Background Art

[0002] In recent years, with the growth of data volume and the improvement of computing power, many new artificial intelligence methods have emerged, including reinforcement learning, transfer learning, and deep learning. Reinforcement learning can operate with limited knowledge of the environment and limited feedback on the quality of decision-making. It is widely used in smart grid networks, communication networks, etc. Transfer learning can migrate the model of big data to small data and realize personalized migration, which is the direction of future development. Deep learning shows the best performance in image processing, such as image classification, object detection, face recognition, and super-resolution. The reason is that through deep learning, the network automatically learns various information features of the image and saves them in the hidden layer.

[0003] The nonlinearity of the infrared focal plane array (FPA) responsivity is a major problem in evaluating infrared images, which will affect the detection and classification capabilities of infrared sensors. Due to the influence of factors such as detector materials and related manufacturing technologies, the responsivity of each detection element to the infrared focal plane array (IRFPA) is different, resulting in IRFPA non-uniformity correction (NUC) being an important technology of IRFPA. With the rapid development of infrared focal plane array (IRFPA) technology, higher requirements are put forward for non-uniformity correction (NUC) technology. If the non-uniformity of infrared images cannot be effectively corrected, it is difficult to separate the target signal from the background. NUC algorithms are mainly divided into two categories: reference-based algorithms and scene-based algorithms; reference-based algorithms rely on standard radiation sources to obtain accurate radiation energy, and the common radiation source is black body. The algorithm is simple and easy to use, and has a good correction effect.

[0004] The infrared focal plane array is the core component of the infrared thermal imaging system and the key to detecting, identifying and analyzing the infrared information of objects. Its performance directly determines the final performance of the infrared thermal imaging system. Ideally, when the infrared focal plane array receives uniform infrared radiation, the radiation response of different detection elements is theoretically consistent. In practical applications, due to the influence of factors such as material manufacturing process, when each pixel of the infrared focal plane array detector receives the same infrared radiation, there is a difference in the intensity of the response output electrical signal (i.e., non-uniformity), and this difference is much higher than that of the visible light sensor. This non-uniform response is usually manifested as a fixed pattern superimposed on the scene image on the infrared image, so it is also called fixed pattern noise. This noise causes the non-uniformity of the camera, and non-uniformity is a special non-uniformity noise. This inherent defect has a great impact on the application of infrared imaging systems. It is common in cooled infrared focal plane arrays and non-cooled frozen infrared focal plane arrays, which seriously affects the quality of infrared images and limits the accuracy of subsequent image processing. Summary of the invention

[0005] In order to give full play to and apply the huge advantages of deep learning in image processing, it is necessary to solve the problem that non-uniform noise seriously affects the quality of infrared images and limits the accuracy of subsequent image processing.

[0006] The present invention aims to solve the problem of column fringes and random noise caused by non-uniform noise in infrared imaging systems, and provides a deep learning non-uniform correction algorithm combined with a reference radiation source in combination with a deep learning method, thereby performing non-uniform correction.

[0007] To this end, the present invention provides a deep learning non-uniformity correction method combined with a reference radiation source, comprising the following steps:

[0008] Step 1: Dataset creation;

[0009] Step 2: Pass the prepared data set through the network model, train the network model in a supervised manner, and determine the network parameters;

[0010] Step 3: Use the trained network model to perform non-uniform correction on the image to be corrected to obtain the corrected image.

[0011] Furthermore, the specific process of step 1, data set preparation is:

[0012] Step 101, prepare a real black body data set, collect black body image data of the infrared camera under working conditions in different temperature ranges, where the temperature range is changed to 5°C to 10°C, and a total of M temperature ranges need to be photographed;

[0013] Step 102: Prepare a black body data set for comparison, using the captured black body image Y pTo obtain the black body image Y for comparison and correction q ;

[0014] Step 103, data set production. In the deep learning process, in order to avoid the need to divide and reorganize the paired data sets when the data sets are produced, the specific operations are as follows:

[0015] First, the paired images at M temperatures are divided into blocks, and the size and position of the paired data set are kept consistent during the block division, where the blocks can be l*s blocks;

[0016] Secondly, the l*s*M*N data set after block division is reorganized to ensure that the size and position of the paired data set remain consistent during block division, and the size of the reorganized image is still m*n;

[0017] At this point, M pairs of blackbody data at different temperatures can be obtained.

[0018] Furthermore, in step 102, the black body data set is prepared by taking the black body image Y p To obtain the black body image Y for comparison and correction q The specific process is:

[0019] First, after obtaining M×N blackbody images, the image pixel data Y at each temperature is calculated. p :

[0020]

[0021] in, That is, it is the effective image of blackbody data acquisition at temperature T1, (i, j) represents the position of each pixel in the image of size m×n, represents the kth blackbody image at temperature T1; thus, M temperatures {T1, T2, …, T M}M bold image data

[0022] Secondly, the paired image Y of the effective black body image is calculated using the black body image data q ;

[0023]

[0024] in, is the contrast-corrected blackbody image at temperature T1, F (i , j) is the non-uniform correction function; then we can get M effective images of blackbody data at a wide temperature. and the corresponding uniform response image

[0025] In the above, F (i , j) The non-uniform correction function can be one-point correction, two-point correction, multi-point correction or a calibration algorithm based on curve fitting.

[0026] Furthermore, in step 2, the prepared data set is passed through the network model, and the network model is trained in a supervised manner to determine the specific process of network parameters:

[0027] Step 201, the black body image is used as the input image x, and firstly, n 3×3 convolutions are used to extract the shallow features F0 of the input image x, as shown in formula (3), and the input image is mapped to a high-dimensional feature space, where:

[0028] F0=f conv (x) (3)

[0029] Step 202: The information is passed through a deep feature extraction module f consisting of three STL Block (STLB) modules, a 3×3 convolutional layer and a residual structure. DF1 , for F0(F0∈R H×W×C ) extracts deep features, as shown in formula (4):

[0030] F DF1 =f DF1 (F0) = f STLB (f STLB (f STLB (f conv (F0)))) (4)

[0031] The STLB module consists of 6 STL layers, a 3×3 convolution and residual structure. Each STL consists of 2 normalization layers (Layer Norm), 1 multi-head self-attention (MSA) module, 1 multi-layer perceptron (MLP) and 2 residual connections.

[0032] The MLP consists of two fully connected layers and a GeLU activation function; the input of size H×W×C is split into M×M local windows and reshaped into is the total number of windows, each window The standard self-attention is calculated at each window as shown in equations (5) and (6);

[0033] Q = XP Q ,K=XP K ,V=XP V (5)

[0034]

[0035] Among them, PQ, PK and PV are projection matrices shared across different windows, B indicates that relative position encoding can be learned. The addition of relative position encoding here can further improve the performance of the model.

[0036] Step 203: Image feature F DF1 After the down sampling block, which includes a 2x down sampling layer and a 3×3 convolution layer, the image feature after the down sampling block is F DSF1 ; Through the deep feature extraction module F consisting of 2 STLB modules, a 3×3 convolutional layer and a residual structure DF2 F DSF1 Extract image features as shown in formula (7):

[0037]

[0038] in, It is a shallow feature. For deep features;

[0039] Image feature F DF2 The image features after the downsampling module are: The image feature size is then restored through the upsampling block; thus, the feature information is completely extracted n times, and the shallow and deep features are reconstructed to obtain the final correction result;

[0040] Step 204, the loss function is used to measure the gap between the predicted result and the expected result. The loss value is continuously input into the network for back propagation to give the network positive and negative feedback to achieve continuous optimization and updating of the network model parameters; the structural similarity index (SSIM) is a full-reference objective quality evaluation indicator for measuring structural similarity, which jointly compares brightness, contrast and structure, so as to obtain the final parameter result by back propagation updating the parameters.

[0041] Furthermore, the specific process of step 3, using the trained network model to perform non-uniform correction on the image to be corrected to obtain the corrected image is: saving the parameter results obtained in step 2 as a weight file xx.pth. By using the weight file to initialize the neural network, the input image to be corrected is input into the network structure to obtain the image after non-uniformity correction.

[0042] The advantages of the present invention are: the present invention provides such a deep learning non-uniformity correction method combined with a reference radiation source, which can give full play to and apply the great advantages of deep learning in image processing, solve the problem that non-uniform noise seriously affects the quality of infrared images and limits the accuracy of subsequent image processing, and deeply mines the image structure features, image detail features, and spatial hierarchy features of blackbody images at different scales by using a deep network, retains all effective image information except non-uniform noise, and realizes high-quality non-uniformity correction of blackbody images.

[0043] The present invention is described in detail below with reference to the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 It is a flow chart of the overall steps of the present invention.

[0045] Figure 2 It is a flow chart of the algorithm network structure of the present invention.

[0046] Figure 3 This is a flow chart of the STLB module in the network of the present invention.

[0047] Figure 4 It is a flow chart of the STL module in the network of the present invention.

[0048] Figure 5 It is a flow chart of the MSA module in the network of the present invention.

[0049] Figure 6 Schematic diagram of the collected blackbody image.

[0050] Figure 7 Schematic diagram of the corrected blackbody image.

[0051] Figure 8 Schematic diagram of the captured scene image.

[0052] Fig. 9 Schematic diagram of the scene image taken after correction. DETAILED DESCRIPTION

[0053] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose, the specific implementation methods, structural features and effects of the present invention are described in detail below in conjunction with the accompanying drawings and embodiments.

[0054] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0055] All features disclosed in this specification, or steps in all methods or processes disclosed, except mutually exclusive features and / or steps, can be combined in any manner.

[0056] Any feature disclosed in this specification (including any additional claims, abstract and drawings), unless otherwise stated, may be replaced by other equivalent or alternative features with similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

[0057] Example 1

[0058] The infrared focal plane array is the core component of the infrared thermal imaging system. It is the key to detecting, identifying and analyzing the infrared information of objects. Its performance directly determines the final performance of the infrared thermal imaging system. Ideally, when the infrared focal plane array receives uniform infrared radiation, the radiation response of different detection elements is theoretically consistent. In practical applications, due to the influence of factors such as material manufacturing process, when each pixel of the infrared focal plane array detector receives the same infrared radiation, there is a difference in the intensity of the response output electrical signal (i.e., non-uniformity), and this difference is much higher than that of the visible light sensor. This non-uniform response is usually manifested as a fixed pattern superimposed on the scene image on the infrared image, so it is also called fixed pattern noise. This noise causes the non-uniformity of the camera, and non-uniformity is a special non-uniformity noise. This inherent defect has a great impact on the application of infrared imaging systems. It is common in cooled infrared focal plane arrays and non-cooled frozen infrared focal plane arrays, which seriously affects the quality of infrared images and limits the accuracy of subsequent image processing.

[0059] In order to give full play to and apply the great advantages of deep learning in image processing, and solve the problem that non-uniform noise seriously affects the quality of infrared images and limits the accuracy of subsequent image processing, this embodiment provides a deep learning non-uniform correction algorithm combined with a reference radiation source to perform non-uniform correction.

[0060] This embodiment provides a Figure 1 to Figure 5 The deep learning non-uniformity correction method combined with a reference radiation source is shown.

[0061] Since the black body image does not contain any scene information, the purest noise information can be extracted. Therefore, the method of the present invention uses the black body data obtained by the radiation source as a training set, and trains through this training set. This network structure is a network structure based on a residual network, which includes shallow feature extraction, multi-scale deep feature extraction and residual structure. First, the black body image is sent to the correction network. The shallow feature extraction module includes a simple convolution layer for preliminary feature extraction of the black body image and extracting the low-frequency information of the input image. Then, the shallow features of different image scales are deeply mined through the residual structure and the multi-scale deep feature extraction module. The high-frequency information of the input image is obtained by using the deep feature extraction module, and the residual connection is matched. Finally, the shallow features of the black body image are spliced ​​with the deep feature channel dimension and input into the image module which is essentially a simple convolution layer. As a whole, the image structure features, image detail features, and spatial hierarchical features of the black body image at different scales are deeply mined by using a deep network, and all effective image information except non-uniform noise is retained to achieve high-quality non-uniformity correction of the black body image.

[0062] In order to give full play to and apply the huge advantages of deep learning in image processing, it is necessary to solve the problem that non-uniform noise seriously affects the quality of infrared images and limits the accuracy of subsequent image processing.

[0063] The present invention aims to solve the problem of column fringes and random noise caused by non-uniform noise in infrared imaging systems, and provides a deep learning non-uniform correction algorithm combined with a reference radiation source in combination with a deep learning method, thereby performing non-uniform correction.

[0064] To this end, the present invention provides a deep learning non-uniformity correction method combined with a reference radiation source, comprising the following steps:

[0065] Step 1: Dataset creation;

[0066] Step 2: Pass the prepared data set through the network model, train the network model in a supervised manner, and determine the network parameters;

[0067] Step 3: Use the trained network model to perform non-uniform correction on the image to be corrected to obtain the corrected image.

[0068] Furthermore, the specific process of step 1, data set preparation is:

[0069] Step 101, prepare a real black body data set, collect black body image data of the infrared camera under working conditions in different temperature ranges, where the temperature range is changed to 5°C to 10°C, and a total of M temperature ranges need to be photographed;

[0070] Take T1 temperature 0℃ as an example. First, adjust the black body to T1 temperature, aim the infrared detector to be calibrated at the black body and take N images, where each image collected has m×n pixels. Then raise the black body temperature to T2 temperature 5℃, and get T i N blackbody images at different temperatures. Repeat the above image acquisition steps to complete the M temperature segments.

[0071] Step 102: Prepare a black body data set for comparison, using the captured black body image Y p To obtain the black body image Y for comparison and correction q ;

[0072] First, after obtaining M×N blackbody images, the image pixel data Y at each temperature is calculated. p , here we take the black body image at temperature T1 as an example:

[0073]

[0074] in, That is, it is the effective image of blackbody data acquisition at temperature T1, (i, j) represents the position of each pixel in the image of size m×n, represents the kth blackbody image at temperature T1; thus, M temperatures {T1, T2, …, T M}M bold images

[0075] Secondly, the paired image Y of the effective black body image is calculated using the black body image data q ;

[0076]

[0077] in, is the contrast-corrected blackbody image at temperature T1, F (i , j) is the non-uniform correction function; then we can get M effective images of blackbody data at a wide temperature. and the corresponding uniform response image

[0078] In the above, F (i , j) The non-uniform correction function can be one-point correction, two-point correction, multi-point correction or a calibration algorithm based on curve fitting.

[0079] Step 103, data set production. In the deep learning process, in order to avoid the need to divide and reorganize the paired data sets when the data sets are produced, the specific operations are as follows:

[0080] First, the paired images at M temperatures are divided into blocks, and the size and position of the paired data set are kept consistent during the block division, where the blocks can be l*s blocks;

[0081] Secondly, the l*s*M*N data set after block division is reorganized to ensure that the size and position of the paired data set remain consistent during block division, and the size of the reorganized image is still m*n;

[0082] At this point, M pairs of blackbody data at different temperatures can be obtained.

[0083] Furthermore, in step 2, the prepared data set is processed as follows: Figure 2 The network model shown uses a supervised approach to train the network model. The specific process of determining the network parameters is:

[0084] Step 201, the black body image is used as the input image x, and firstly, n 3×3 convolutions are used to extract the shallow features F0 of the input image x, as shown in formula (3), and the input image is mapped to a high-dimensional feature space, where:

[0085] F0=f conv (x) (3)

[0086] Among them, R represents a real number set, H represents the height of the image, and W represents the height and width C of the image. in Indicates the number of input channels, f conv is the convolution function;

[0087] Step 202: The information is passed through a deep feature extraction module f consisting of three STL Block (STLB) modules, a 3×3 convolutional layer and a residual structure. DF1 , for F0(F0∈R H×W×C ) extracts deep features, as shown in formula (4):

[0088] F DF1 =f DF1 (F0) = f STLB (f STLB (f STLB (f conv (F0)))) (4)

[0089] The STLB module consists of 6 STL layers, a 3×3 convolution and residual structure. For each submodule, the STLB module structure is as follows: Figure 3 Each STL consists of 2 normalization layers (Layer Norm), 1 multi-head self-attention (MSA) module, 1 multi-layer perceptron (MLP) and 2 residual connections; the STL module structure is shown in Figure 4 shown.

[0090] MLP contains 2 fully connected layers and GeLU activation function; MSA module is as follows Figure 5 As shown, the input of size H×W×C is split into local windows of size M×M and reshaped as is the total number of windows, each window The standard self-attention is calculated at each window as shown in equations (5) and (6);

[0091] Q = XP Q ,K=XP K ,V=XP V (5)

[0092]

[0093] Among them, C represents the number of channels, d represents the dimension of the vector, and P Q , P K and P V is the projection matrix shared across different windows, B indicates that relative position encoding can be learned. The addition of relative position encoding here can further improve the performance of the model.

[0094] Step 203: Image feature F DF1 After the down sampling block, which includes a 2x down sampling layer and a 3×3 convolution layer, the image feature after the down sampling block is F DSF1 ; Through the deep feature extraction module f consisting of 2 STLB modules, a 3×3 convolutional layer and a residual structure DF2 F DSF1 Extract image features as shown in formula (7):

[0095]

[0096] in, It is a shallow feature. For deep features;

[0097] Image feature F DF2 The image features after the downsampling module are: The image feature size is then restored through the upsampling block; thus, the feature information is completely extracted n times, and the shallow and deep features are reconstructed to obtain the final correction result;

[0098] Step 204, the loss function is used to measure the gap between the predicted result and the expected result. The loss value is continuously input into the network for back propagation to give the network positive and negative feedback to achieve continuous optimization and updating of the network model parameters; the structural similarity index (SSIM) is a full-reference objective quality evaluation indicator for measuring structural similarity, which jointly compares brightness, contrast and structure, so as to obtain the final parameter result by back propagation updating the parameters.

[0099] Furthermore, the specific process of step 3, using the trained network model to perform non-uniform correction on the image to be corrected to obtain the corrected image is: saving the parameter results obtained in step 2 as a weight file xx.pth. By using the weight file to initialize the neural network, the input image to be corrected is input into the network structure to obtain the image after non-uniformity correction.

[0100] In summary, this embodiment provides a deep learning non-uniformity correction method combined with a reference radiation source, which can give full play to and apply the great advantages of deep learning in image processing, solve the problem that non-uniform noise seriously affects the quality of infrared images and limits the accuracy of subsequent image processing, and deeply mine the image structure features, image detail features, and spatial hierarchy features of blackbody images at different scales by using a deep network, retain all valid image information except non-uniform noise, and achieve high-quality non-uniformity correction of blackbody images.

[0101] Example 2

[0102] The captured scene image is corrected using the deep learning non-uniform correction method combined with the reference radiation source shown in Example 1. First, black body images are collected at different temperatures, such as Figure 6 The diagram shows a schematic diagram of a collected black body image. The collected black body image is corrected to construct a paired data set, as shown in FIG. Figure 7 Schematic diagram of the corrected blackbody image.

[0103] The proposed network is trained with the paired dataset constructed above. The trained network model is used to correct the captured scene images, such as Figure 8 As shown in FIG. , it is a schematic diagram of the captured scene image. Fig. 9 As shown, it is a schematic diagram of the scene image captured after correction.

[0104] Depend on Fig. 9 The correction results shown in FIG. 4 show that the deep learning non-uniformity correction method combined with the reference radiation source provided by the present invention has a significant correction effect; all effective image information except the non-uniformity noise can be retained to achieve high-quality non-uniformity correction of the blackbody image.

[0105] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A deep learning non-uniformity correction method combined with a reference radiation source, characterized in that: The steps include: Step 1: Dataset creation; Step 2: Pass the prepared data set through the network model, train the network model in a supervised manner, and determine the network parameters; Step 3: Use the trained network model to perform non-uniform correction on the image to be corrected to obtain the corrected image.

2. The deep learning non-uniformity correction method combined with a reference radiation source as claimed in claim 1, characterized in that: The specific process of step 1, data set preparation is: Step 101, prepare a real black body data set, collect black body image data of the infrared camera under working conditions in different temperature ranges, where the temperature range is changed to 5°C to 10°C, and a total of M temperature ranges need to be photographed; Step 102: Prepare a black body data set for comparison, using the captured black body image Y p To obtain the black body image Y for comparison and correction q ; Step 103, data set production. In the deep learning process, in order to avoid the need to divide and reorganize the paired data sets when the data sets are produced, the specific operations are as follows: First, the paired images at M temperatures are divided into blocks, and the size and position of the paired data set are kept consistent during the block division, where the blocks can be l*s blocks; Secondly, the l*s*M*N data set after block division is reorganized to ensure that the size and position of the paired data set remain consistent during block division, and the size of the reorganized image is still m*n; At this point, M pairs of blackbody data at different temperatures can be obtained.

3. The deep learning non-uniformity correction method combined with a reference radiation source as claimed in claim 2, characterized in that: In step 102, the black body data set is prepared by taking a black body image Y p To obtain the black body image Y for comparison and correction q The specific process is: First, after obtaining M×N blackbody images, the image pixel data Y at each temperature is calculated. p : in, That is, it is the effective image of blackbody data acquisition at temperature T1, (i, j) represents the position of each pixel in the image of size m×n, represents the kth blackbody image at temperature T1; thus, M temperatures {T1, T2, …, T M }M bold image data Secondly, the paired image Y of the effective black body image is calculated using the black body image data q ; in, is the contrast-corrected blackbody image at temperature T1, F (i，j) is the non-uniform correction function; then we can get M effective images of blackbody data at a wide temperature. and the corresponding uniform response image In the above, F (i , j) The non-uniform correction function can be one-point correction, two-point correction, multi-point correction or a calibration algorithm based on curve fitting.

4. The deep learning non-uniformity correction method combined with a reference radiation source as claimed in claim 1, characterized in that: In step 2, the prepared data set is passed through the network model, and the network model is trained in a supervised manner to determine the specific process of network parameters: Step 201, the black body image is used as the input image x, and firstly, n 3×3 convolutions are used to extract the shallow features F0 of the input image x, as shown in formula (3), and the input image is mapped to a high-dimensional feature space, where: F0=f conv (x) (3) Step 202: The information is passed through a deep feature extraction module f consisting of three STL Block (STLB) modules, a 3×3 convolutional layer and a residual structure. DF1 , for F0(F0∈R H×W×C ) extracts deep features, as shown in formula (4): F DF1 =f DF1 (F0)=f STLB (f STLB (f STLB (f conv (F0)))) (4) The STLB module consists of 6 STL layers, a 3×3 convolution and residual structure. Each STL consists of 2 normalization layers (Layer Norm), 1 multi-head self-attention (MSA) module, 1 multi-layer perceptron (MLP) and 2 residual connections. The MLP consists of two fully connected layers and a GeLU activation function; the input of size H×W×C is split into M×M local windows and reshaped into is the total number of windows, each window The standard self-attention is calculated at each window as shown in equations (5) and (6); Q=XP Q ,K=XP K ,V=XP V (5) Among them, PQ, PK and PV are projection matrices shared across different windows, B indicates that relative position encoding can be learned. The addition of relative position encoding here can further improve the performance of the model. Step 203: Image feature F DF1 After the down sampling block, which includes a 2x down sampling layer and a 3×3 convolution layer, the image feature after the down sampling block is F DSF1 ; Through the deep feature extraction module f consisting of 2 STLB modules, a 3×3 convolutional layer and a residual structure DF2 F DSF1 Extract image features as shown in formula (7): in, It is a shallow feature. For deep features; Image feature F DF2 The image features after the downsampling module are: Then the image feature size is restored through the up sampling module (Up samplingBlock); thus, the feature information is completely extracted n times, and the shallow and deep features are reconstructed to obtain the final correction result; Step 204, the loss function is used to measure the gap between the predicted result and the expected result. The loss value is continuously input into the network for back propagation to give the network positive and negative feedback to achieve continuous optimization and updating of the network model parameters; the structural similarity index (SSIM) is a full-reference objective quality evaluation indicator for measuring structural similarity, which jointly compares brightness, contrast and structure, so as to obtain the final parameter result by back propagation updating the parameters.

5. The deep learning non-uniformity correction method combined with a reference radiation source as claimed in claim 1, characterized in that: The specific process of step 3, using the trained network model to perform non-uniform correction on the image to be corrected to obtain the corrected image, is as follows: the parameter results obtained in step 2 are saved as a weight file xx.pth. By using the weight file to initialize the neural network, the image to be corrected is input into the network structure to obtain the image after non-uniformity correction.

Images