Infrared thermal radiation effect bias field correction method based on parameterized polynomial surface

Abstract

The application discloses an infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface, and belongs to the technical field of image processing, and comprises the following steps: infrared thermal radiation effect degradation mechanism modeling, constructing a thermal radiation effect bias field degradation model based on parameterized polynomial surface fitting; simulation training dataset construction, including paired infrared image data before and after degradation; knowledge-guided parameterized polynomial network construction, including a first network unit, a second network unit, a bias field image reconstruction unit and a loss calculation unit. The degraded image to be corrected is input into the first network unit of the trained infrared thermal radiation effect bias field correction model, and a restored clear infrared image is output. The application utilizes prior knowledge of polynomial fitting bias field for modeling, and guides the network to perform self-adaptive correction, thereby increasing the interpretability of the network, and effectively improving the generalization performance of the method in a real measurement scene and the imaging quality of an image.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and more particularly relates to an infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface. BACKGROUND

[0002] Infrared imaging has the advantages of high detection accuracy, good concealment and strong anti-interference ability, and is one of the important means for precise guidance of high-speed aircrafts. Aerodynamic optics is an interdisciplinary subject combining aerodynamics and optics, and is also an important direction to support the development of aircrafts. Among them, aerodynamic optical thermal radiation correction is an important application direction of the subject in the field of aerospace.

[0003] The aerodynamic thermal radiation effect of an infrared image mainly comes from the radiation of a high-temperature optical window. When a high-speed aircraft carrying an infrared imaging system flies in the atmosphere, due to aerodynamic heating, the imaging system is in a serious aerodynamic thermal environment, at the same time, the optical head cover interacts with the high-speed airflow, inducing an aerodynamic thermal radiation effect similar to "halo", which saturates the imaging and degrades the signal-to-noise ratio of the target image, seriously restricting the imaging quality and detection accuracy of the optical imaging detection system.

[0004] At present, the existing aerodynamic thermal radiation effect correction methods mainly fall into two categories: the first category is a traditional method based on an optimization model, and the second category is a learning method based on data driving.

[0005] The method based on the optimization model driving converts the degraded image from an image subspace to other subspaces by artificially designing a hand-crafted prior, utilizes the discriminative features between the image and the bias field in the subspace, and solves the optimization iteration by constructing an energy functional to effectively separate the image and the bias field. This category of methods can directly process a single image without paired image pairs, and has high generalization for real scenes. However, the hand-crafted prior knowledge is mostly based on a simplified mechanism model, and in the actual optimization solving process, a linear transformation is mainly used, so the feature expression ability is poor.

[0006] The learning method based on data driving relies on the strong feature expression ability of a nonlinear network, and directly fits the complex characteristics of the degradation effect through a large amount of data learning, but this method has poor interpretability. In actual scenes, it is difficult to obtain image pairs, although existing technologies can obtain more real bias field data pairs before and after degradation through wind tunnels, rocket jacks and other experiments, but the cost of obtaining is high, the scene is simple, and the complex spatiotemporal variation characteristics of the infrared bias field during imaging of a real high-speed aircraft cannot be fully simulated. Therefore, a large amount of simulation data is needed to supplement the training data set.

[0007] In summary, the current technology has the following problems: the first type of method does not have strong feature expression capability, and the performance is limited in actual correction; the second type of method has poor generalization ability and can only have good performance in specific scenarios. SUMMARY

[0008] In view of the defects of the prior art and the need for improvement, the present application provides an infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface, which aims to model the bias field prior knowledge and guide the network to adaptively correct it, so that it is more interpretable, and at the same time, the network's powerful feature expression capability is used to learn the differences between different structures of images and bias fields, solving the problems of poor imaging quality of infrared images due to thermal radiation effect and limited correction effect of the current technology.

[0009] To achieve the above-mentioned purpose, according to one aspect of the present application, an infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface is provided, comprising:

[0010] Training stage: taking the degraded image Y as a training sample and the clear image corresponding to the degraded image as a label, a infrared thermal radiation effect bias field correction model is trained; wherein the infrared thermal radiation effect bias field correction model comprises a first network unit, a second network unit, a bias field image reconstruction unit and a loss calculation unit;

[0011] The first network unit is used to predict the clear infrared image in the degraded image Y;

[0012] The second network unit is used to predict the bias field parameters in the degraded image Y;

[0013] The bias field image reconstruction unit is used to obtain the predicted bias field image by polynomial surface fitting based on the predicted bias field parameters;

[0014] The loss calculation unit is used to minimize the consistency loss of the predicted clear infrared image and the input clear infrared image C as the target, to calculate the training loss with the predicted clear infrared image and the predicted bias field image regularization prior information as the constraint, and to adjust the parameters of the first network unit and the second network unit in the reverse direction to make the loss converge, so that the first network unit outputs the restored clear infrared image;

[0015] Application stage: input the degraded image to be corrected into the first network unit of the trained infrared thermal radiation effect bias field correction model, and output the restored clear infrared image.

[0016] Further, the degraded image Y is obtained by the following way:

[0017] Construct a thermal radiation effect bias field image B;

[0018] The thermal radiation effect bias field image B is superimposed on the clear infrared image C to obtain the degradation image Y, Y=B+C.

[0019] Further, the process of constructing the thermal radiation effect bias field image B comprises:

[0020] Step S11, determining the bias field order D and bias field parameter a of the bias field image t,s , wherein 0≤t≤D, 0≤s≤D-t, max(t+s)=D, bias field parameter a t,s number

[0021] Step S12, based on the bias field parameter a t,s , the bias field image is fitted by a polynomial surface to obtain the fitted thermal radiation effect bias field image B.

[0022] Further, the polynomial surface fitting formula is:

[0023]

[0024] Wherein, B represents the thermal radiation effect bias field image, (x i ,y j ) is the thermal radiation effect bias field image coordinate, D is the bias field order, a t,s is the bias field parameter.

[0025] Further, in the loss calculation unit, the loss function for calculating the training loss is:

[0026]

[0027] Wherein, λ and μ are regularization parameters, a′ t,s represent the predicted bias field parameter, R represents the predicted clear infrared image, and the subscripts i and j represent the coordinates (i, j).

[0028] Further, the first network unit comprises a first multi-scale encoder with a residual module and a multi-scale decoder connected in series;

[0029] The degradation image Y is input into the first network unit, and after feature extraction by the first multi-scale encoder with a residual module, it is input into the multi-scale decoder, and the predicted clear image is output.

[0030] Further, the second network unit comprises a second multi-scale encoder with a residual module and a fully connected layer connected in series;

[0031] The degraded image Y is input into a second network unit, and after feature extraction by a second multi-scale encoder of the residual module, is input into a full connection layer to output the predicted bias field parameter a' t,s .

[0032] Further, the feature extraction process of the multi-scale encoder includes:

[0033] The degraded image Y is input, and the clear infrared image C is taken as a label, and through multiple convolution operations and corresponding down-sampling operations, a feature map is output.

[0034] The multi-scale encoder is a first multi-scale encoder or a second multi-scale encoder.

[0035] Further, the process of outputting the predicted clear image by the first multi-scale decoder includes:

[0036] The feature map output by the first multi-scale encoder is subjected to an up-sampling operation, and the feature map obtained by up-sampling is combined with a feature map of a corresponding size obtained by down-sampling in a channel to obtain a predicted clear image; wherein the scale of the up-sampling operation is the same as the scale of the corresponding down-sampling operation.

[0037] According to a second aspect of the present application, a computer storage medium is provided, comprising instructions which, when executed on a computer, cause the computer to perform the infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface as any one of the first aspect.

[0038] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0039] (1) The infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface can establish an infrared bias field degradation model, learn the bias field degradation characteristics under the unified optimization framework of the model, reconstruct the bias field image through the infrared thermal radiation effect bias field correction model based on the polynomial surface fitting, guide the network correction using prior knowledge, learn the discriminative features between the degradation and the image from the degraded data, form a spatial dimension discriminative expression, fully separate the two, and better remove the bias field in the infrared image to obtain a high-quality clear image, thereby solving the technical problems of poor imaging quality of infrared images due to thermal radiation effects and limited correction effect of current technologies, and improving the imaging quality.

[0040] (2) The application adopts a data-driven mode to learn the characteristics of clear infrared images and bias field images, and the speed in the test stage is much faster than that of a traditional optimization model-driven method; in the process of learning the characteristics of the bias field images, the bias field images are fitted by a polynomial surface, and the accuracy and interpretability are higher than those of a learning method using only a network, so that the infrared bias field images under the effect of aerodynamic heat radiation can have a good restoration effect.

[0041] (3) In the application, according to the infrared bias field degradation model Y=C+B, a large number of thermal radiation effect bias field images are constructed and superimposed on the clear images to obtain corresponding simulation degradation images, so that the amount of the data set is increased. The inherent properties of the images are mined by using a large amount of simulation data, so that the accuracy of model training can be improved. Moreover, considering that the bias field is the thermal radiation generated by the aerodynamic heat effect when the aircraft flies at a high speed, which is superimposed on the original infrared radiation to cause a degradation effect, the infrared bias field degradation model constructed in the application is more reasonable than the multiplicative model in the prior art, and the simulation difficulty is also reduced.

[0042] In summary, the infrared thermal radiation effect bias field correction method based on the parameterized polynomial surface can guide the network to perform adaptive correction by using the prior knowledge of the polynomial fitting bias field, increase the interpretability of the network, also has the powerful feature expression ability of the deep learning method, can effectively improve the generalization performance of the method in the actual measurement scene and the imaging quality of the infrared image, and provides a new idea for the infrared image aerodynamic thermal radiation effect bias field correction. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a whole process schematic diagram of the infrared thermal radiation effect bias field correction method based on the parameterized polynomial surface provided by the embodiment of the application.

[0044] Figure 2 is a principle diagram of the infrared thermal radiation effect bias field correction method based on the parameterized polynomial surface provided by the embodiment of the application.

[0045] Figure 3 is a structure schematic diagram of the first network unit and the second network unit provided by the embodiment of the application.

[0046] Figure 4 is a 16-bit infrared clear image and a corresponding 8-bit image provided by the embodiment of the application, wherein (a) and (b) in the figure respectively represent the 16-bit infrared clear image and the corresponding 8-bit image.

[0047] Figure 5are the recovery results of the simulated infrared bias field images provided by the embodiments of the present application, wherein (a) and (d) in the figure are original clear infrared images, (b) and (e) are the simulated degraded images, and (c) and (f) are the corresponding recovery results.

[0048] Figure 6 are the recovery results of the actual infrared bias field images provided by the embodiments of the present application, wherein (a)-(c) in the figure are different degraded images in the actual scene, and (d)-(f) are the corresponding recovery results. DETAILED DESCRIPTION

[0049] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0050] In the present application, the terms "first", "second", etc. in the present application and the drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence.

[0051] As shown in Figures 1-3 The present application provides an infrared thermal radiation effect bias field correction method based on a parameterized polynomial surface, which includes a training stage and an application stage.

[0052] The training stage: taking the degraded image Y as a training sample and taking the clear image corresponding to the degraded image as a label, the infrared thermal radiation effect bias field correction model is trained; wherein the infrared thermal radiation effect bias field correction model mainly includes: a first network unit, a second network unit, a bias field image reconstruction unit and a loss calculation unit.

[0053] Specifically, the constructed thermal radiation effect bias field is used to construct the degraded infrared image, wherein the constructed degraded image is used as the data set for training the first network unit and the second network unit. Specifically, by constructing the thermal radiation effect bias field image B, the infrared bias field degradation model Y=C+B is established, and the constructed thermal radiation effect bias field image B is superimposed on the clear infrared image C to obtain the degraded image Y, wherein C is the clear infrared image.

[0054] The first network unit is used to take the degraded image Y as input and output the predicted clear infrared image R;

[0055] The second network unit is used to take the degraded image Y as input and output the predicted bias field parameter;

[0056] a bias field image reconstruction unit configured to obtain a predicted bias field image B' by polynomial surface fitting based on the predicted bias field parameters;

[0057] a loss calculation unit configured to train the first network unit and the second network unit by minimizing a consistency loss between the predicted clear infrared image R and the input clear infrared image C, and by using the predicted clear infrared image R and the predicted bias field image B' as regularization prior information, to calculate a loss, to update parameters in the first network unit and the second network unit by backpropagation of a network gradient, and to make the entire model converge until a training round reaches a set upper limit or the loss is less than a set threshold, so that the first network unit outputs a restored clear infrared image.

[0058] Specifically, the process of constructing the thermal radiation effect bias field image B includes:

[0059] Step S11, determining a bias field order D of the bias field image, a number num of bias field parameters, and corresponding bias field parameters a t,s , where 0≤t≤D, 0≤s≤D-t, and max(t+s)=D, and num represents the number of bias field parameters a t,s .

[0060] In this embodiment, D is 4, and num=10; the bias field parameters a t,s are random numbers in [0, 1], and in the example, a t,s takes the value [0, 0.012, 0.009, 0.661, 0.044, 0.009, 0.001, 0.087, 0.004, 0.001].

[0061] Step S12, fitting the bias field image based on the bias field parameters a t,s to obtain a fitted thermal radiation effect bias field image B; where the fitting formula is:

[0062]

[0063] where B is the fitted thermal radiation effect bias field image, (x i ,y j ) is the thermal radiation effect bias field image coordinate, D is the bias field order, and a t,s is the bias field parameter.

[0064] Specifically, in step S11, the bias field order D and the bias field parameters a t,s of the bias field image constructed according to the existing thermal radiation effect bias field image are estimated.

[0065] Specifically, the first network unit and the second network unit share a sample input end, take the degraded image Y as a training sample, take the clear infrared image as a label, and are respectively input into the first network unit and the second network unit to simultaneously train the two network units.

[0066] The first network unit comprises a first multi-scale encoder with a residual module and a multi-scale decoder connected in series, the degraded image Y is input into the first network unit, feature extraction is performed on the degraded image Y by the first multi-scale encoder with a residual module, and then the multi-scale decoder is used to output a predicted clear image R.

[0067] The process of feature extraction by the first multi-scale encoder with a residual module comprises: taking the degraded image Y as an input and the clear infrared image C as a label, performing multiple convolution and corresponding down-sampling operations, and outputting a feature map.

[0068] The process of outputting the predicted clear image R by the multi-scale decoder comprises: performing an up-sampling operation on the feature map output by the encoder, performing channel merging on the feature map obtained by up-sampling and a corresponding down-sampled feature map of the same size, and obtaining the predicted clear image R; wherein the scale of the up-sampling operation is the same as the scale of the corresponding down-sampling.

[0069] In this embodiment, the process of feature extraction by the first multi-scale encoder with a residual module and the process of outputting the predicted clear image R by the multi-scale decoder specifically comprise:

[0070] The input single-channel degraded image extracts features by 16 3x3 convolutions and one ReLU function, changes the number of image channels to 16, and does not change the image size, and then performs one global pooling to reduce the image size to

[0071] Features are extracted by 32 3x3 convolutions and one ReLU function, the number of image channels is changed to 32, the image size is not changed, and then one global pooling is performed to reduce the image size to

[0072] Features are extracted by 64 3x3 convolutions and one ReLU function, the number of image channels is changed to 64, the image size is not changed, and then one global pooling is performed to reduce the image size to

[0073] By analogy, five convolutions and corresponding down-sampling operations are performed to change the number of image channels to 256 and the image size to

[0074] Then, the multi-scale decoder is further used:

[0075] The 256-dimensional feature map is processed by 128 3x3 convolutions to change the image channel number to 128 and keep the image size unchanged, and then is up-sampled by interpolation to expand the image size to 4 times of the feature map size;

[0076] The 128-dimensional feature map of the same size obtained in the corresponding down-sampling process is combined with the feature map in the channel dimension to obtain a 256-dimensional feature map. The 256-dimensional feature map is processed by 128 3x3 convolutions to change the image channel number to 128 and keep the image size unchanged. The 128-dimensional feature map is processed by 64 3x3 convolutions to change the image channel number to 64 and keep the image size unchanged, and then is up-sampled by interpolation to expand the image size to 16 times of the original feature map size. The 64-dimensional feature map of the same size obtained in the corresponding down-sampling process is combined with the feature map in the channel dimension to obtain a 128-dimensional feature map. The 128-dimensional feature map is processed by 64 3x3 convolutions to change the image channel number to 64 and keep the image size unchanged. In this way, a 16-dimensional feature map of the same size as the original image is finally obtained. Through a 3x3 convolution and a ReLU function, a predicted clear infrared image is obtained, and the predicted clear infrared image is constrained by regularized prior information:

[0077] Φ(R) = ∑ i,j (|R i+1,j -R i,j |+|R i,j+1 -R i,j |)。

[0078] Specifically, the second network unit includes a second multi-scale encoder with a residual module connected in series and a fully connected layer. The degraded image Y is input into the second network unit, the feature of the degraded image Y is extracted by the second multi-scale encoder with a residual module, and the predicted bias field parameter a t,s is output by the fully connected layer.

[0079] The second multi-scale encoder for extracting the bias field feature is the same as the first multi-scale encoder, and finally a 256-dimensional feature image is obtained. The predicted bias field parameter a t,s is obtained by a fully connected layer.

[0080] The predicted bias field image B' output by the bias field image reconstruction unit is:

[0081]

[0082] wherein a′ t,s represents the predicted bias field parameter of the second network unit.

[0083] In the loss calculation unit, the regularization prior constraint of the predicted clear infrared image R is:

[0084]

[0085] wherein, R i,j represents a predicted clear infrared image, and subscripts i and j represent coordinates (i, j).

[0086] The prior constraint of the predicted bias field image B' regularization is:

[0087]

[0088] wherein, B'(x i ,y j ) represents a predicted bias field image B', a' t,s represents a predicted bias field parameter.

[0089] The loss calculation formula provided by the embodiment of the present application is:

[0090]

[0091] In the loss formula, the first term is a data term conforming to the degradation mechanism, the second term is a clear image prior knowledge modeling, the third term is a bias field degradation prior knowledge modeling, and λ and μ are regularization parameters used to balance the relationship between the terms.

[0092] In the application stage, the to-be-corrected degraded image is input into the first network unit in the trained infrared thermal radiation effect bias field correction model, and the restored clear infrared image is output.

[0093] The present application will be further described below in combination with a simulation experiment.

[0094] (1) Constructing simulation data

[0095] The simulation degraded data is constructed according to the bias field degradation model. The construction process consists of three steps: first, the gray scale range of the 16-bit infrared clear image is scaled to the 8-bit image with a range of [0, 255], as shown in (a) and (b) of FIG. 1; second, using Matlab, the fitting formula of the bias field image is fitted according to the polynomial surface, the bias field is calculated through the for loop, and is superimposed on the 16-bit clear image to obtain the 16-bit degraded image; third, the 16-bit degraded image is displayed as an 8-bit image to facilitate viewing of the restoration effect. Figure 4

[0096] (2) Constructing a degraded data set

[0097] 500 16-bit infrared images are selected, and are uniformly cropped to 640x512 size, the maximum order is selected to be 4, 20 different bias fields are simulated for each picture, a total of 10000 simulated bias field images are generated, the original clear infrared image is used as the corresponding label to constitute the data set. The data set is randomly divided into a training set and a test set, and the ratio of the training set and the test set is 8:2. ​

[0098] (3) Train the first network unit and the second network unit

[0099] Set all uninitialized parameters in the network to 0. Scale the degraded image and its corresponding sharp image using the same standard: subtract the minimum value of the sharp image from each image and divide by the difference between the maximum value of the degraded image and the minimum value of the sharp image, ensuring that both scaled images are within the range [0,1]. Input the processed degraded image and labels into the network to be trained and calculate the total loss (Loss). Use network gradient backpropagation to feed the total loss into the network to update the parameters, and repeat the above steps continuously, adjusting the learning rate to decrease the total loss. If the number of repetitions reaches the upper limit or the total loss is less than a certain threshold, stop training. The final result is used as the model parameters after network training, for subsequent practical applications of the model.

[0100] (4) Test

[0101] The 16-bit degraded images of the test set and measured data are scaled to 640×512 and then input into the network to obtain the output image and bias field. The output image is the restored clear infrared image.

[0102] (5) Performance Evaluation

[0103] Peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) are used to evaluate the quality of the algorithm's restoration results.

[0104] like Figure 5 As shown in (c) and (f), these are the results of the simulated infrared bias field image restoration. Figure 5 In the diagram, (a) and (d) are the original clear infrared images, (b) and (e) are the degraded images obtained through simulation, i.e., the degraded images obtained by superimposing the constructed thermal radiation effect bias field image B onto the clear infrared image C, and (c) and (f) are the corresponding restored results. It should be noted that the images shown here are the results of converting 16-bit original images to 8-bit images for display. The method of this invention is faster than traditional methods, processing each image in approximately 0.03 seconds; and compared to existing network-based methods, it better preserves image details.

[0105] Figure 6 In the figure, (d)-(f) represent the actual infrared bias field image restoration results. Figures (a)-(c) show different degraded images in the actual scene, and (d)-(f) represent the corresponding restoration results. It should be noted that the images shown here are all 16-bit original images converted to 8-bit images for display. The restoration results demonstrate that the method of this invention can be generalized to actual measured images.

[0106] The present application, by establishing an infrared bias field degradation model, under the unified optimization framework of the model, decouples the bias field and the clear image by using the pre-constructed multi-scale network, learns the bias field degradation characteristics through the network, obtains the predicted bias field parameters, reconstructs the bias field image by using the bias field model based on the parametric polynomial surface fitting, and uses the image prior as the training loss to guide the network correction, can better remove the bias field in the infrared image, obtain a high-quality clear image, and solve the technical problems of poor imaging quality of the infrared image due to the thermal radiation effect and limited correction effect of the current technology, and improve the imaging quality. Since the characteristics of the clear infrared image and the bias field image are learned by the network under the unified infrared bias field degradation model, the speed in the test stage is much faster than that of the traditional optimization model driven method; meanwhile, in the process of learning the characteristics of the bias field image, the bias field image is fitted by using the polynomial, and the accuracy and interpretability are higher than those of the learning method using the network alone, and the infrared bias field image under the effect of aerodynamic thermal radiation can have a good restoration effect, and meet the requirements of good and fast correction of the image under the interference of thermal radiation in the field of aviation and aerospace.

[0107] At the same time, the present application simultaneously utilizes the advantages of the optimization model and the neural network, and provides a new idea for the bias field correction of the infrared image aerodynamic thermal radiation effect.

[0108] Those skilled in the art can easily understand that the above description is only a preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. An infrared thermal radiation effect bias field correction method based on a parameterized polynomial curved surface, characterized in that, The method comprises the following steps: The training phase: with the degraded image The degraded image corresponding to the clear image is taken as a label, and the infrared thermal radiation effect bias field correction model is trained; wherein the infrared thermal radiation effect bias field correction model comprises: a first network unit, a second network unit, a bias field image reconstruction unit and a loss calculation unit; the first network element is configured to predict a clear infrared image in the degraded image the first network element is configured to predict a clear infrared image in the degraded image The second network element is configured to predict a bias field parameter in the degraded image ; The bias field image reconstruction unit is used to obtain a predicted bias field image by polynomial surface fitting based on the predicted bias field parameters; The loss calculation unit is used to minimize the predicted sharp infrared image and the input sharp infrared image. The training loss is calculated with the consistency loss as the objective and the regularized prior information of the predicted clear infrared image and the predicted bias field image as constraints. The parameters of the first network unit and the second network unit are adjusted in reverse to make the loss converge, so that the first network unit outputs the restored clear infrared image. The application stage comprises the following steps: The degraded image to be corrected is input into the first network unit of the trained infrared thermal radiation effect bias field correction model, and a restored clear infrared image is output; The loss function for calculating the training loss is as follows: wherein and is a regularization parameter, denotes the predicted bias field parameters, R denotes the predicted clear infrared image, the subscript denotes the coordinates ); is the order of the bias field, is the thermal radiation effect bias field image coordinate.

2. The correction method of claim 1, wherein the degraded image by the following way: A thermal radiation effect bias field image B is constructed; superimposing the thermal radiation effect bias field image B on the sharp infrared image above, to obtain the degraded image , .

3. The correction method of claim 2, wherein The process of constructing the thermal radiation effect bias field image B comprises the following steps: Step S11, determining the bias field order of the bias field image and bias field parameters wherein , , bias field parameters number of ; Step S12, based on bias field parameters The bias field image B is obtained by using a polynomial surface to fit the bias field image.

4. The correction method of claim 3, wherein The polynomial surface fitting formula is as follows: where B represents a thermal radiation effect bias field image, is a thermal radiation effect bias field image coordinate, is a bias field order, is a bias field parameter.

5. The correction method of claim 1, wherein The first network unit comprises a first multi-scale encoder with a residual module and a multi-scale decoder connected in series; The degraded image Y is input into the first network unit, is subjected to feature extraction by the first multi-scale encoder with a residual module, is input into the multi-scale decoder, and a predicted clear image is output.

6. The correction method of claim 5, wherein, The second network unit comprises a second multi-scale encoder with a residual module and a full connection layer connected in series; The degraded image Y is input into a second network unit, and after feature extraction by a second multi-scale encoder of the residual module, is input into the full connection layer, and the predicted bias field parameter is output .

7. The correction method of claim 6, wherein, The process of feature extraction by the multi-scale encoder comprises the following steps: with the degraded image as input, output a feature map through multiple convolution operations and corresponding down-sampling operations; The multi-scale encoder is the first multi-scale encoder or the second multi-scale encoder.

8. The correction method of claim 7, wherein, The process of outputting the predicted clear image by the multi-scale decoder comprises the following steps: The feature map output by the first multi-scale encoder is subjected to an upsampling operation, the feature map obtained by the upsampling operation is combined with a feature map obtained by downsampling in a channel, and a predicted clear image is obtained; wherein the scale of the upsampling operation is the same as the scale of the corresponding downsampling operation.

9. A computer storage medium comprising instructions, the computer storage medium being characterized by: The instructions make the computer execute the infrared thermal radiation effect bias field correction method based on the parameterized polynomial surface when the instructions run on the computer.

Images