Long-wave infrared image aberration correction method based on intensity mapping and deep learning

Abstract

The invention discloses a long-wave infrared image aberration correction method based on intensity mapping and deep learning, and the method comprises the steps: dividing a clear long-wave infrared image into a plurality of intensity intervals according to the intensity range of the clear long-wave infrared image, dividing the wavelength range of long-wave infrared into a plurality of wavelength intervals, and mapping the intensity intervals to the wavelength intervals; using ray tracing software to extract point spread functions of central wavelengths of different wavelength intervals of the optical system; carrying out convolution on the clear long-wave infrared pictures in different intensity intervals and the mapped point spread functions to obtain aberration degraded long-wave infrared pictures; weighting the aberration degenerated long-wave infrared picture according to the normalized light intensity response of the wavelength response curve of the detector corresponding to the central wavelength mapped by the aberration degenerated long-wave infrared picture, and adding noise to obtain a simulated aberration degenerated long-wave infrared picture; and enabling the clear pictures and the simulation pictures to be in one-to-one correspondence, constructing a data set, and training a deep learning aberration correction network. According to the invention, the aberration simulation precision and the aberration correction capability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of optical measurement and wavefront detection, and particularly relates to a long-wave infrared image aberration correction method based on intensity mapping and deep learning. BACKGROUND

[0002] Long-wave infrared refers to electromagnetic waves with a wavelength in the range of 8-14 microns, which is exactly in one of the Earth's atmospheric windows, with the characteristics of strong penetration of haze and small interference from solar radiation. A long-wave infrared camera generates images by detecting the thermal radiation emitted by the surface of an object, and its core physical basis is the theory of thermal radiation. According to the Stefan-Boltzmann law, the total radiation power of an object is proportional to the fourth power of its absolute temperature, which can be expressed as:

[0003]

[0004] where, is the radiation power per unit area, is the emissivity (between 0 and 1), is the Stefan-Boltzmann constant, is the absolute temperature of the object. This means that the higher the temperature, the stronger the energy emitted by the object, and the stronger the intensity signal received by the infrared camera, which is represented as higher intensity (brighter or whiter) on the image. In order to accurately measure the temperature, the infrared camera needs to be radiometrically calibrated. Usually, the camera will calibrate a blackbody with a known temperature to establish the correspondence between intensity and temperature, that is, linear correction. In this way, within a reasonable temperature range, the intensity and temperature are approximately linearly related, that is, the higher the intensity, the higher the temperature. According to the Wien displacement law, temperature is inversely proportional to the wavelength of radiation. The output intensity can be approximately expressed as:

[0005]

[0006] where, is the correction coefficient, is the wavelength, is the background noise offset. Therefore, in the model we established, the local intensity of the long-wave infrared picture is approximately inversely proportional to the received wavelength.

[0007] However, like images captured in other wavelengths, long-wave infrared (LWIR) images are subject to degradation in the lens imaging system. This degradation primarily stems from optical aberrations, including spherical aberration, coma, and astigmatism. These aberrations cause the point spread function to expand, blurring image edges and losing detail. This, combined with inherent noise in infrared detectors, further degrades image quality. Therefore, LWIR image restoration has become a key technology for improving infrared image quality, enhancing target recognition, and improving temperature measurement accuracy. Traditional image restoration methods, such as deconvolution, can improve image quality to a certain extent, but these methods often rely on accurate imaging models or degradation kernels, resulting in poor robustness in real-world scenarios. In recent years, deep learning-based image restoration methods have made significant progress. Their core idea is to train on a large number of real or simulated "clear-degraded" image pairs to learn the mapping relationship for recovering high-quality images from degraded images. Compared to traditional methods, deep learning can fit complex nonlinear degradation processes and exhibits stronger generalization capabilities, making it particularly suitable for processing infrared images of complex scenes. However, existing methods do not take into account the characteristic that the intensity of long-wave infrared images is directly related to the wavelength of radiation, resulting in low accuracy of aberration simulation and affecting the quality of the final image restoration. Summary of the Invention

[0008] In response to the shortcomings of the existing technology, the present invention proposes a long-wave infrared image aberration correction method based on intensity mapping and deep learning. According to the characteristics of long-wave infrared aberration, a mapping relationship between its intensity and wavelength is established. For image components with different intensities, different central wavelength point spread functions are used to simulate and weight the aberration, thereby improving the accuracy of aberration simulation. The aberration simulation dataset is used to train a deep learning network, so that the deep learning network can perform aberration correction and restoration on degraded infrared images, thereby obtaining the final clear imaging long-wave infrared image.

[0009] The purpose of the present invention is achieved through the following technical solutions:

[0010] A long-wave infrared image aberration correction method based on intensity mapping and deep learning includes the following steps:

[0011] S1: Divide the clear long-wave infrared image into several intensity intervals according to its own intensity range, divide the wavelength range of the typical long-wave infrared into several wavelength intervals, and map the intensity intervals to the corresponding wavelength intervals;

[0012] S2: Use ray tracing software to extract the point spread function of the central wavelength of different wavelength ranges of the long-wave infrared optical system;

[0013] S3: convolve the clear long-wave infrared picture of different intensity intervals with the point spread function of the mapped center wavelength to obtain aberration-degraded long-wave infrared pictures of different intensity intervals;

[0014] S4: weight the aberration-degraded long-wave infrared pictures of different intensity intervals according to the normalized light intensity response of the wavelength response curve of the detector corresponding to the mapped center wavelength, and add noise to obtain simulated aberration-degraded long-wave infrared pictures;

[0015] S5: one-to-one correspondence between the clear long-wave infrared picture and the simulated aberration-degraded long-wave infrared picture, construct a training data set, and train a deep learning aberration correction network; the input of the deep learning aberration correction network is an aberration-degraded long-wave infrared picture, and the output is a clear long-wave infrared picture;

[0016] S6: collect a new long-wave infrared aberration-degraded picture, input the trained deep learning aberration correction network for aberration correction.

[0017] Further, the S1 is specifically implemented by the following sub-steps:

[0018] S1.1: first normalize the intensity range of the clear long-wave infrared picture itself, then divide the normalized intensity range into N intensity intervals according to the intensity from small to large, and then store the clear long-wave infrared picture according to the intensity intervals as N new pictures;

[0019] S1.2: divide the wavelength range [8um, 12um] of the typical long-wave infrared design into N wavelength intervals according to the wavelength from long to short, and calculate the center wavelength of the i-th wavelength interval i ;

[0020] S1.3: establish an intensity mapping model, one-to-one correspondence between N intensity intervals and N wavelength intervals, establish a linear mapping relationship, and the higher the intensity, the shorter the wavelength, the lower the intensity, the longer the wavelength.

[0021] Further, in the S2, for each center wavelength, first divide the image plane of the long-wave infrared optical system into U fields of view along the horizontal direction and V fields of view along the vertical direction, and calculate the point spread function of each field of view.

[0022] Further, in the S3, along the horizontal direction, divide the N new pictures stored according to the intensity intervals into U fields of view and V fields of view;

[0023] Then, convolve each intensity image block after field division with the corresponding center wavelength point spread function, and perform convolution operation on each field to obtain a predicted image block;

[0024] Finally, for any central wavelength, the image is divided into U fields of view horizontally and V fields of view vertically. The predicted image blocks are spliced ​​together to obtain a spliced ​​image of the same size as the original clear long-wave infrared image, which is used as the aberration-degraded long-wave infrared image of the intensity range corresponding to the central wavelength.

[0025] Furthermore, the deep learning aberration correction network includes an intensity decomposition module, a correction layer, a splicing layer and an improved U-shaped network connected in sequence;

[0026] The intensity decomposition module is used to divide the intensity into N different intensity intervals according to the intensity distribution of the input image, and map the input image into N feature maps;

[0027] The correction layer includes N correction blocks, and the correction blocks are connected to the feature Figure One One-to-one correspondence is used to make targeted preliminary corrections to the degradation characteristics of different intensity intervals;

[0028] The splicing layer is used to splice the N feature maps corrected by the N correction blocks into a new N-channel fused feature map to complete the preliminary restoration of intensity perception;

[0029] The downsampling path and upsampling path of the improved U-shaped network both include four stages. Each stage of the downsampling path uses one of the correction blocks for feature extraction, followed by 2×2 maximum pooling downsampling, gradually compressing the spatial dimension and expanding the number of channels to capture multi-scale contextual information. The encoder finally outputs features that are transmitted to the decoder via an intermediate connection. The upsampling path also includes four stages. Each stage first performs upsampling through a 2×2 transposed convolution, and then splices with the high-resolution features transmitted by the corresponding level of the encoder through a jump connection to achieve detail reuse. The splicing result is further refined and repaired by a correction block, gradually restoring the spatial resolution and improving the fidelity of details. Finally, a 1×1 convolution layer is used to map the high-dimensional features to the final image space to generate a clear long-wave infrared image after aberration correction, completing the end-to-end aberration correction process.

[0030] Furthermore, all correction blocks in the deep learning aberration correction network have the same structure, and the correction blocks perform three-way parallel processing on the input features to achieve separation and extraction of multi-scale degradation information, wherein:

[0031] The first path compresses the number of channels through 1×1 convolution, enhances nonlinearity through ReLU activation function, and further fuses features through 1×1 convolution, focusing on perceiving the detail degradation pattern of local areas in the image;

[0032] The second path uses a 1×1 convolution and ReLU followed by a 5×5 convolution layer to expand the receptive field and extract a large range of low-frequency aberration components and background distortion information in the image;

[0033] The third path captures aberration structures and context features in a medium scale range through 1x1 convolution, ReLU and 3x3 convolution operations;

[0034] Then, the three output features are spliced in the channel dimension, and then ReLU activation and 1x1 convolution are performed to realize cross-channel feature fusion and dimension reduction, and finally the feature information after the correction block repair is output.

[0035] Further, the correction layer is not less than two layers, and each layer includes N correction blocks.

[0036] A long-wave infrared image aberration correction device based on intensity mapping and deep learning includes one or more processors for implementing a long-wave infrared image aberration correction method based on intensity mapping and deep learning.

[0037] An electronic device includes:

[0038] One or more processors;

[0039] A storage device for storing one or more programs, when the one or more programs are executed by the electronic device, the electronic device implements a long-wave infrared image aberration correction method based on intensity mapping and deep learning.

[0040] A computer-readable storage medium having a program stored thereon, the program being executed by a processor to implement a long-wave infrared image aberration correction method based on intensity mapping and deep learning.

[0041] The beneficial effects of the present application are as follows:

[0042] Unlike the traditional method of directly simulating aberration through a gray-scale picture, the method of the present application maps intensity values to different center wavelengths based on the monochromaticity of long-wave infrared radiation and the direct relationship between intensity and wavelength, thereby using different point spread functions to convolve the components in different intensity intervals of the picture, thereby simulating the aberration degradation of the long-wave infrared picture. After constructing the aberration data set and the deep learning network, the network learns the aberration correction process, thereby effectively correcting the actual long-wave infrared aberration degradation picture. The present application improves the aberration simulation accuracy and aberration correction ability. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The flowchart of the long-wave infrared image aberration correction method based on intensity mapping and deep learning of the present application.

[0044] Figure 2 The structure diagram of the deep learning aberration correction network.

[0045] Figure 3Schematic diagram of the structure of the correction block, the basic component unit of the network.

[0046] Figure 4 These are the aberration correction results of the actual target image, where (a) is the actual long-wave infrared optical structure diagram, (b) is the blurred target image obtained from the actual shot, and (c) is the clear target image output by the aberration correction network.

[0047] Figure 5 Schematic diagram of the results of aberration correction for long-wave infrared natural images. (a) is the first infrared natural image with aberration, (b) is the first clear natural image output by the aberration correction network, (c) is the second infrared natural image with aberration, and (d) is the second clear natural image output by the aberration correction network. DETAILED DESCRIPTION

[0048] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments, and the purpose and effects of the present invention will become more apparent. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0049] like Figure 1 As shown, a long-wave infrared image aberration correction method based on intensity mapping and deep learning in one embodiment of the present invention includes the following steps:

[0050] Step 1: Divide the clear long-wave infrared image into several intensity intervals according to its own intensity range, divide the wavelength range of typical long-wave infrared into several wavelength intervals, and map the intensity intervals to the corresponding wavelength intervals.

[0051] Step 1 includes the following sub-steps:

[0052] S1.1: First, normalize the intensity range of the clear long-wave infrared image itself, then divide the normalized intensity range into N intensity intervals from small to large, and then store the original clear long-wave infrared image P into N new images according to the intensity intervals; where the i-th new image is represented as P(I i );

[0053] Among them, the i-th normalized intensity interval I i satisfy:

[0054]

[0055] S1.2: Divide the wavelength range of a typical long-wave infrared design [8μm, 12μm] into N wavelength intervals from long to short wavelengths. The wavelength range covered by the i-th wavelength interval is W i for:

[0056]

[0057] the center wavelength λ of the i-th wavelength interval i is:

[0058]

[0059] Considering a typical long-wave infrared night shooting scene, the higher the temperature of an object, the stronger the intensity it presents in the resulting picture, thus helping to discover the target, while its corresponding wavelength is shorter. Thus, the higher the intensity of the shot, the shorter the wavelength, and the lower the intensity, the longer the wavelength, so the mapping relationship between intensity and wavelength is approximated as a linear mapping, and an intensity mapping model is established.

[0060] S1.3: Establish an intensity mapping model to one-to-one correspond N intensity intervals with N wavelength intervals, establish a linear mapping relationship, and the higher the intensity, the shorter the wavelength, and the lower the intensity, the longer the wavelength.

[0061] At this time, according to the established mapping relationship, the new picture P(I i ) can also be written as P(λ i ).

[0062] Step two: Use ray tracing software to extract the point spread function of the center wavelength of different wavelength intervals of the long-wave infrared optical system.

[0063] Because the aberration is spatially unevenly distributed, it is preferred to first divide the image plane into U fields of view horizontally and V fields of view vertically, and the total extracted point spread function matrix expression is

[0064]

[0065] The corresponding clear picture P(λ i ) is also divided into U fields of view horizontally and V fields of view vertically, and the expression of the clear picture of each field of view is P(u,v,λ i ), so the number of clear image blocks is one-to-one corresponding to the point spread function matrix.

[0066] Step three: Convolve the clear long-wave infrared picture of different intensity intervals with the point spread function of the mapped center wavelength to obtain the aberration-degraded long-wave infrared picture of different intensity intervals;

[0067] Each intensity image block after field division is respectively convolved with the corresponding center wavelength point spread function, and the predicted image block is obtained, and its expression is

[0068]

[0069] wherein, represent a convolution operation.

[0070] for any one central wavelength λ i , the predicted image block is spliced back into a spliced image with the same size as the original clear long-wave infrared picture, and the i-th central wavelength λ i corresponds to the spliced image The expression is:

[0071]

[0072] where Patch represents a splicing function.

[0073] the spliced image is the aberration-degraded long-wave infrared picture of the i-th intensity interval.

[0074] Step four: the aberration-degraded long-wave infrared pictures of different intensity intervals are weighted according to the normalized light intensity response of the wavelength response curve of the corresponding detector of the mapped central wavelength, and noise is added to obtain a simulated aberration-degraded long-wave infrared picture.

[0075] The expression of the simulated aberration-degraded long-wave infrared picture is:

[0076]

[0077] where, is the normalized light intensity response of the detector to the i-th central wavelength λ i , n is the added noise, which conforms to a Gaussian distribution , wherein represents a Gaussian distribution, and is the standard deviation of the Gaussian distribution.

[0078] Step five: the clear long-wave infrared picture and the simulated aberration-degraded long-wave infrared picture are corresponded one by one to construct a training data set, and a deep learning aberration correction network is trained. The input of the deep learning aberration correction network is an aberration-degraded long-wave infrared picture, and the output is a clear long-wave infrared picture.

[0079] As shown in Figure 2 , the deep learning aberration correction network takes a correction block with a multi-branch feature extraction structure as the core basic unit, and includes a strength decomposition module, a plurality of correction layers, a splicing layer and an improved U-net network connected in sequence. The strength decomposition module divides the intensity into N different intensity intervals according to the intensity distribution of the input image, and maps the input image into N feature maps, which are respectively sent into N correction blocks of the first correction layer for independent processing. The correction block and the feature Figure OneOne-to-one correspondence is used for targeted preliminary correction of different intensity interval degradation characteristics. In this embodiment, the correction layer is two layers, and the output of the first layer correction layer is one-to-one corresponding to the input of the second layer correction layer for processing. After the processing of the two layers of correction layers in succession, the N feature maps after preliminary correction are spliced again by the splicing layer to combine into a new N-channel fusion feature map, and the preliminary restoration of the intensity perception is completed.

[0080] The new N-channel fusion feature map is further input into the improved U-net network for deep fusion and detail restoration. The down-sampling path and the up-sampling path of the improved U-shaped network each include four stages. Each stage of the down-sampling path uses a correction block for feature extraction, followed by 2x2 maximum pooling down-sampling, gradually compressing the spatial dimension and expanding the number of channels to capture multi-scale context information; the final output feature of the encoder is transmitted to the decoder through intermediate connection; the up-sampling path also includes four stages, each of which first performs up-sampling through 2x2 transpose convolution, and then splices the high-resolution features transmitted by the corresponding level of the encoder through the jump connection to realize detail reuse; the splicing result is further refined and repaired by a correction block, gradually restoring the spatial resolution and improving the detail fidelity; finally, a 1x1 convolution layer is used to map the high-dimensional features to the final image space to generate a clear long-wave infrared image after aberration correction, completing the end-to-end aberration correction process.

[0081] As shown in Figure 3 , in view of the characteristics of long-wave infrared images in noise distribution, detail blur and low-frequency aberration, three parallel processing is implemented on the input features in each correction block to realize the separation and extraction of multi-scale degradation information: the first path compresses the number of channels through 1x1 convolution, enhances the nonlinearity through ReLU activation function, and further fuses the features through 1x1 convolution to focus on the perception of the detail degradation mode in the local area of the image; the second path expands the receptive field through 1x1 convolution and ReLU followed by a 5x5 convolution layer to extract the large-scale low-frequency aberration component and background distortion information in the image; the third path captures the aberration structure and context features in the medium scale range through 1x1 convolution, ReLU and 3x3 convolution operation; then, the output features of the three paths are spliced in the channel dimension, and ReLU activation and 1x1 convolution are performed to realize cross-channel feature fusion and dimension reduction, and finally the feature information repaired by the correction block is output.

[0082] The loss function of the deep learning aberration correction network is

[0083]

[0084] The gradient descent method is used to train the deep learning aberration correction network to learn the aberration correction process.

[0085] Step six: collect the actual photos taken by the long-wave infrared optical system, input the trained deep learning aberration correction network, and obtain the clear picture after aberration correction.

[0086] Corresponding to the foregoing embodiments of the long-wave infrared image aberration correction method based on intensity mapping and deep learning, the application also provides embodiments of a long-wave infrared image aberration correction device based on intensity mapping and deep learning.

[0087] The long-wave infrared image aberration correction device based on intensity mapping and deep learning provided by the embodiments of the application comprises one or more processors for implementing the long-wave infrared image aberration correction method based on intensity mapping and deep learning in the foregoing embodiments.

[0088] The embodiments of the long-wave infrared image aberration correction device based on intensity mapping and deep learning of the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiments can be implemented by software, or by hardware or a combination of software and hardware. Taking software implementation as an example, as a logically meaningful device, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory for execution by the processor of the device with data processing capability where it is located. From the hardware level, in addition to the processor, memory, network interface, and non-volatile memory, the device with data processing capability where the embodiments are located usually includes other hardware according to the actual functions of the device with data processing capability, and details are not repeated here.

[0089] The implementation process of the functions and roles of each unit in the above device is specifically described in the implementation process of the corresponding steps in the above method, and is not repeated here.

[0090] For the device embodiments, since they basically correspond to the method embodiments, the related parts are described in the part of the method embodiments. The device embodiments described above are only illustrative, and the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place or distributed on multiple network units. According to actual needs, part or all of the modules can be selected to achieve the purpose of the application scheme. Those skilled in the art can understand and implement without creative labor.

[0091] The embodiments of the application also provide a computer readable storage medium having a program stored thereon, which, when executed by a processor, implements the long-wave infrared image aberration correction method based on intensity mapping and deep learning in the foregoing embodiments.

[0092] The computer readable storage medium can be an internal storage unit of any of the aforementioned data processing capable devices, such as a hard disk or a memory. The computer readable storage medium can also be an external storage device, such as a plug-in hard disk, a SmartMedia card (SMC), an SD card, a Flash card, etc. equipped on the device. Further, the computer readable storage medium can include both an internal storage unit of any of the aforementioned data processing capable devices and an external storage device. The computer readable storage medium is used to store the computer program and other programs and data required by the aforementioned data processing capable devices, and can also be used to temporarily store data that has been output or will be output.

[0093] A specific application example of the method of the present application is given below to illustrate the technical effects of the present application.

[0094] S1: 1000 clear infrared pictures of 640x512 size are divided into 10 intensity intervals, and the same wavelength range [8μm, 12μm] is also divided into 10 wavelength intervals according to wavelength from long to short, and the intensity intervals and the wavelength intervals form a one-to-one mapping relationship.

[0095] S2: The long-wave infrared optical system contains a single aspherical germanium lens, and a diffraction surface is processed on the front surface, and the image plane is divided into 10x8 fields of view, each field of view is 64x64 in size, and the point spread function of the center wavelength of each field of view is extracted using the optical design software ZEMAX, and the center wavelengths are 11.8μm, 11.4μm, 11μm, 10.6μm, 10.2μm, 9.8μm, 9.4μm, 9.0μm, 8.6μm, and 8.2μm.

[0096] S3: The clear image blocks of 10 center wavelengths of 10x8 fields of view are respectively convolved with the point spread function to obtain aberration-degraded long-wave infrared pictures of different intensity intervals.

[0097] S4: The aberration-degraded pictures of different intensity intervals are weighted according to the normalized intensity [0.05, 0.08, 0.1, 0.12, 0.15, 0.15, 0.12, 0.1, 0.08, 0.05] of the corresponding wavelength response curve, and Gaussian noise with a standard deviation of 0.02 is added to obtain the final simulated aberration-degraded long-wave infrared picture.

[0098] S5: The obtained clear and aberration-degraded long-wave infrared images are mapped one by one to form a data set, and a deep learning aberration correction network is trained. Its input is the aberration-degraded long-wave infrared image, and the output target is a clear long-wave infrared image. The learning rate is set to 0.001, and the back-propagation Adam algorithm is used to implement gradient descent. It is iterated and optimized 300 times to obtain the trained deep learning aberration correction network.

[0099] S6: Take a real infrared image or a blurred long-wave infrared image with aberrations, and use the trained deep learning aberration correction network to process it to obtain a clear image.

[0100] like Figure 4 The figure shows the aberration correction result of the real-shot target image, where (a) is the structure diagram of the optical system, (b) is the result of direct imaging of the target through the optical system, and (c) is the restoration result using the method of the present invention. Figure 4 As can be seen in (b), there is a large phase difference in the target result of direct imaging, but after restoration using the method of the present invention, the target is clear, which proves that the present invention can achieve effective aberration simulation and correction of long-wave infrared optical systems.

[0101] like Figure 5 As shown in Figure 2, a blurred long-wave infrared image containing aberrations is fed into a trained deep learning aberration correction network to produce a clear long-wave infrared image. As can be seen from the figure, the details of the image are clearly restored.

[0102] Those skilled in the art will understand that the foregoing descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art will still be able to modify the technical solutions described in the foregoing examples or substitute equivalents for some of the technical features therein. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the invention shall be included within the scope of protection of the invention.

Claims

1. A long-wave infrared image aberration correction method based on intensity mapping and deep learning, characterized in that: The following steps are involved: S1: Divide the clear long-wave infrared image into several intensity intervals according to its own intensity range, divide the wavelength range of the typical long-wave infrared into several wavelength intervals, and map the intensity intervals to the corresponding wavelength intervals; S2: Use ray tracing software to extract the point spread function of the central wavelength of different wavelength ranges of the long-wave infrared optical system; S3: convolve the clear long-wave infrared images of different intensity ranges with the point spread function of their mapped central wavelength to obtain aberration-degraded long-wave infrared images of different intensity ranges; S4: The aberration-degraded long-wave infrared images of different intensity ranges are weighted according to the normalized light intensity response of the wavelength response curve of the detector corresponding to the mapped central wavelength, and noise is added to obtain a simulated aberration-degraded long-wave infrared image; S5: The clear long-wave infrared image and the simulated aberration-degraded long-wave infrared image are matched one by one, a training data set is constructed, and a deep learning aberration correction network is trained; the input of the deep learning aberration correction network is the aberration-degraded long-wave infrared image, and the output is a clear long-wave infrared image; S6: Collect new long-wave infrared aberration-degraded images and input them into the trained deep learning aberration correction network for aberration correction.

2. The long-wave infrared image aberration correction method based on intensity mapping and deep learning according to claim 1, characterized in that: The S1 is specifically implemented through the following sub-steps: S1.1: First, normalize the intensity range of the clear long-wave infrared image itself, and then divide the normalized intensity range into N intensity intervals, and then the clear long-wave infrared image is stored according to the intensity intervals as N New pictures; S1.2: Divide the wavelength range of typical long-wave infrared designs [8μm, 12μm] into N wavelength intervals, and calculate the i The central wavelength of the wavelength interval λ i ; S1.3: Establish an intensity mapping model to N Intensity range and N The wavelength ranges are matched one to one to establish a linear mapping relationship, and the higher the intensity, the shorter the wavelength, and the lower the intensity, the longer the wavelength.

3. The long-wave infrared image aberration correction method based on intensity mapping and deep learning according to claim 2, characterized in that: In S2, for each central wavelength, the image plane of the long-wave infrared optical system is first divided into U Field of view, divided vertically into V Fields of view, calculate the point spread function of each field of view.

4. The long-wave infrared image aberration correction method based on intensity mapping and deep learning according to claim 3, characterized in that: In the S3, there are two parts: U Field of view, vertically divided into V First, store the fields according to the intensity range. N Each new picture is divided into image blocks; Then, the image blocks of each intensity after the field of view is divided are convolved with the point spread function of the corresponding central wavelength, field by field, to obtain the predicted image blocks; Finally, for any central wavelength, it is divided into U Field of view, divided vertically into V The predicted image blocks are stitched together to obtain a stitched image of the same size as the original clear long-wave infrared image, which is used as the aberration-degraded long-wave infrared image of the intensity range corresponding to the central wavelength.

5. The long-wave infrared image aberration correction method based on intensity mapping and deep learning according to claim 1, characterized in that: The deep learning aberration correction network includes an intensity decomposition module, a correction layer, a splicing layer and an improved U-shaped network connected in sequence; The intensity decomposition module is used to divide the intensity into N different intensity intervals according to the intensity distribution of the input image, and map the input image into N feature maps; The correction layer includes N correction blocks, each of which corresponds to the feature map one by one and is used to perform targeted preliminary correction on degradation characteristics in different intensity intervals; The splicing layer is used to splice the N feature maps corrected by the N correction blocks into a new N-channel fused feature map to complete the preliminary restoration of intensity perception; The downsampling path and upsampling path of the improved U-shaped network both include four stages. Each stage of the downsampling path uses one of the correction blocks for feature extraction, followed by 2×2 maximum pooling downsampling, gradually compressing the spatial dimension and expanding the number of channels to capture multi-scale contextual information. The final output features of the encoder are passed to the decoder via an intermediate connection; the upsampling path also includes four stages. Each stage first performs upsampling through a 2×2 transposed convolution, and then splices with the high-resolution features transmitted by the corresponding level of the encoder through a jump connection to achieve detail reuse. The splicing result is further refined and repaired by a correction block, gradually restoring the spatial resolution and improving the fidelity of details; finally, a 1×1 convolution layer is used to map the high-dimensional features to the final image space, generating a clear long-wave infrared image after aberration correction, completing the end-to-end aberration correction process.

6. The long-wave infrared image aberration correction method based on intensity mapping and deep learning according to claim 5, characterized in that: All correction blocks in the deep learning aberration correction network have the same structure. The correction blocks perform three-way parallel processing on the input features to achieve separation and extraction of multi-scale degradation information, where: The first path compresses the number of channels through 1×1 convolution, enhances nonlinearity through ReLU activation function, and further fuses features through 1×1 convolution, focusing on perceiving the detail degradation pattern of local areas in the image; The second path uses a 1×1 convolution and ReLU followed by a 5×5 convolution layer to expand the receptive field and extract a large range of low-frequency aberration components and background distortion information in the image; The third path captures the aberration structure and contextual features in the medium scale range through 1×1 convolution, ReLU and 3×3 convolution operations; Afterwards, the three-way output features are spliced ​​in the channel dimension, and then cross-channel feature fusion and dimensionality reduction are achieved through ReLU activation and 1×1 convolution, and finally the feature information of the repaired correction block is output.

7. The long-wave infrared image aberration correction method based on intensity mapping and deep learning according to claim 6, characterized in that: The correction layers are no less than two layers, and each layer includes N correction blocks.

8. A long-wave infrared image aberration correction device based on intensity mapping and deep learning, characterized in that: The method comprises one or more processors for implementing the long-wave infrared image aberration correction method based on intensity mapping and deep learning as described in any one of claims 1 to 7.

9. An electronic device, characterized in that: include: one or more processors; A storage device for storing one or more programs, which, when executed by the electronic device, enables the electronic device to implement the long-wave infrared image aberration correction method based on intensity mapping and deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by a processor, the long-wave infrared image aberration correction method based on intensity mapping and deep learning as described in any one of claims 1 to 7 is implemented.

Images