Infrared image contrast enhancement method and system capable of keeping spatial domain consistency

Abstract

The invention discloses an infrared image contrast enhancement method and an infrared image contrast enhancement system capable of keeping airspace consistency, which are used for solving the problems of low image contrast and fuzzy details caused by factors such as severe weather and environment temperature interference in infrared image shooting. According to the method, spatial information interaction and channel information interaction are utilized to construct an infrared image enhancement model based on spatial-domain consistency maintenance, spatial-domain consistency expression is extracted on low-resolution features, an enhancement curve is adjusted in a feature domain through a pixel-by-pixel curve, detail information is injected through local interaction on multi-scale expression, and a multi-scale image is obtained. A global enhancement loss function, a multi-scale structure loss function and a histogram loss function are used for training. Compared with an existing infrared enhancement method, the method has the advantages that the low-contrast infrared image can be effectively enhanced, the spatial domain consistency and detail quality of infrared image enhancement are effectively improved, the problems of inconsistent pixel intensity, artifacts and detail blurring in a uniform area of an enhancement result are solved, and the visual effect of the enhancement result is improved.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to an infrared image contrast enhancement method and system for maintaining spatial domain consistency. Background Art

[0002] In recent years, infrared image enhancement technology has important application value in the fields of security monitoring, industrial inspection, etc. However, since infrared images are often collected under extreme environmental conditions such as low light, foggy, overexposed, etc., the imaging quality is low, the details are blurred and the contrast is insufficient. For this reason, researchers have proposed many image enhancement technologies to improve image visibility. Traditional image enhancement methods mainly include histogram equalization and methods based on the Retinex theory. The histogram equalization method improves the contrast by stretching the pixel value distribution, but it is easy to generate noise and artifacts, and it is necessary to manually adjust parameters for different scenes, and the generality is poor. The Retinex method assumes that an image can be decomposed into illumination and reflection components, and improves the image quality by enhancing the reflection component. However, this assumption often does not hold under complex illumination conditions, and it is also difficult to effectively introduce regularization constraints.

[0003] With the development of deep learning technology, more and more learning-based enhancement methods have been proposed. These methods can be mainly divided into three categories: end-to-end network structures, network structures based on the Retinex theory, and plug-and-play enhancement modules. Representative end-to-end networks such as LLNet (LLNet: A deep autoencoder approach to natural low-light image enhancement) and LightenNet (Lightennet: A convolutional neural network for weakly illuminated image enhancement) first use convolutional neural networks to directly output enhanced images. However, due to the simple single-stage network structure, it is difficult to achieve fine detail enhancement. Subsequently, some scholars introduced local kernel prediction modules (LEDNet: Joint low-light enhancement and deblurring in the dark) or long-distance modeling modules (Ultra-high definition low-light image enhancement: A benchmark and transformer based method) to further improve image details and global information. However, such methods usually have a high computational complexity, and the models lack interpretability. In addition, although the Retinex deep network represented by RetinexNet (Sparse gradient regularized deep Retinex network for robust low-light image enhancement) has good interpretability, it still cannot effectively handle complex infrared imaging conditions. At the same time, image enhancement methods based on the plug-and-play concept, such as ZeroDCE (Zero-reference deep curve estimation for low-light image enhancement), ChebyLighter (Chebylighter: Optimal curve estimation for low-light image enhancement), adjust the image brightness through parameterized curves. Although these methods are simple and effective, they are mainly for visible light images and lack specific designs for the degradation phenomena unique to infrared images. Therefore, it is difficult to handle well the problems such as noise and inconsistent spatial structure commonly existing in infrared images.

[0004] For the particularity of infrared images, specialized restoration and enhancement methods have also been proposed. Some single-stage networks such as TherSuRNet (TherISuRNet - a computationally efficient thermal image super-resolution network), ChaSNet (Channel split convolutional neural network (ChaSNet) for thermal image super-resolution), etc. attempt to balance performance and computational complexity, while other works such as PSRGAN (Infrared image super-resolution via transfer learning and PSRGAN) handle more challenging conditions through multi-stage networks. However, existing methods mostly focus on overall brightness enhancement, with limited detail performance and insufficient spatial consistency, resulting in blurred and artifact problems in the enhanced infrared images. And this method can effectively enhance low-contrast infrared images, effectively improving the spatial consistency and detail quality of infrared image enhancement, eliminating the problems of inconsistent pixel intensities, artifacts, and blurred details in the uniform regions of the enhancement results, and improving the visual effect of the enhancement results. Summary of the Invention

[0005] Aiming at the technical defects of insufficient spatial consistency and insufficient detail enhancement existing in the existing infrared image enhancement technology, the present invention proposes a method for enhancing the contrast of infrared images while maintaining spatial consistency. By using spatial information interaction and channel information interaction, an infrared image enhancement model based on maintaining spatial consistency is constructed. An expression of spatial consistency is extracted on low-resolution features, and the enhancement curve is adjusted in the feature domain through a pixel-by-pixel curve; detail information is injected through local interaction in multi-scale expressions; a global enhancement loss function, a multi-scale structure loss function, and a histogram loss function are used for training.

[0006] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In the first aspect, the present invention provides a method for enhancing the contrast of infrared images while maintaining spatial consistency, which uses spatial information interaction and channel information interaction to enhance the contrast of infrared images and improve the image detail quality, including the following steps:

[0008] Train an infrared image enhancement model, where the infrared image enhancement model includes a feature extraction subnet, a contrast enhancement subnet, and a detail enhancement subnet;

[0009] The described feature extraction subnet is used to encode a given infrared degraded image with low contrast and detail loss into a low-resolution feature representation. The contrast enhancement subnet is used to enhance the low-resolution feature representation in the feature domain to obtain an enhanced feature representation. The detail enhancement subnet is used to gradually enhance the detail structure of the enhanced feature representation and reconstruct the original resolution image.

[0010] Using the trained infrared image enhancement model, enhance the contrast of the given infrared degraded image and output the enhancement result.

[0011] Furthermore, the described feature extraction subnet adopts a non-linear activation-free network block.

[0012] Furthermore, the contrast enhancement subnet is composed of several cascaded global enhancement modules (GEMs). Each GEM maintains the generation of spatial consistency descriptors through long-range spatial information interaction and cross-channel information interaction, and performs global enhancement with consistency preservation in the feature domain.

[0013] Furthermore, the calculation process of the GEM is as follows:

[0014] Denote the input of the nth GEM as F n ∈R h×w×c , where h, w, and c are the length, width, and number of channels of the input image respectively. Adopt a two-layer self-attention mechanism, that is, the spatial interaction layer and the channel interaction layer respectively implement long-range spatial information interaction and cross-channel information interaction. Among them, the dimension of the attention map of the spatial interaction layer is hw×hw, and the dimension of the attention map of the channel interaction layer is c×c. Then add the outputs of the two layers to obtain the global consistency descriptor P.

[0015] Use a curve generator to generate an enhancement curve tensor C∈R n from the input F h ×w×K of the GEM and the global consistency descriptor P, where K is the number of curve adjustment iterations;

[0016] Use the enhancement curve tensor C to perform iterative curve adjustment on F n , and the result of the last adjustment is used as the output.

[0017] Furthermore, during iterative curve adjustment, the kth curve adjustment uses the kth channel slice in the enhancement curve tensor C as the enhancement curve parameter, denoted as C k ∈R h×w×1 , and perform the following curve adjustment:

[0018]

[0019] where, is the kth feature domain enhancement result, and ⊙ is the element-wise multiplication;

[0020] Furthermore, the detailed enhancement subnet is composed of a number of cascaded local enhancement modules (LEMs). Each LEM consists of a resolution reconstructor and a detail injector, which gradually enhance and reconstruct high-resolution details. The enhanced features of the last LEM are decoded into the enhanced result of the original resolution through a convolutional layer.

[0021] Furthermore, the calculation process of the LEM is as follows:

[0022] Use the resolution reconstructor to preliminarily reconstruct the high-resolution features from the low-resolution enhanced features and denote it as F content , which is reconstructed. The dimension of the obtained F content is the same as that of . The resolution reconstructor is an upsampling convolutional layer, which are the input and output of the m-th LEM respectively;

[0023] The detail injector includes a dual attention layer and a weighted fusion operation. First, in the dual attention layer, F content and F detail serve as the query feature Q, key-value feature K, and value feature V of the neighborhood attention to calculate the detail-aware enhanced feature and the contrast-aware enhanced feature where F detail is the detail feature extracted from the infrared degraded image by a convolutional layer and has the same dimension as F content . Second, generate the fusion weights of the detail-aware enhanced feature and the contrast-aware enhanced feature through a weight generator, and then perform weighted fusion on the two to output the high-resolution enhanced feature

[0024] Furthermore, the loss function during the training of the infrared image enhancement model includes a global enhancement loss function, a local enhancement loss function, and a histogram loss function;

[0025] The global enhancement loss function L GE is:

[0026]

[0027] where represents the enhanced result generated from the low-resolution feature representation, represents a convolutional decoding layer, and Y ↓M represents downsampling the ground truth image to 1 / 2 M times the original image size to match the size of the low-resolution enhanced result; ‖.‖1 is the L1 norm;

[0028] Set multi-scale features The size is 1 / 2 of the original image M-m , and the local enhancement loss function L LE is as follows:

[0029]

[0030] where M is the number of LEMs, represents the enhancement result generated from the multi-scale features, and Y ↓M represents downsampling the ground truth image to 1 / 2 of the original size m times to match the size of the multi-scale enhancement result, and SSIM represents structural similarity;

[0031] The histogram loss function L hist is as follows:

[0032]

[0033] where h represents the histogram of the enhancement result, represents the histogram of the ground truth image.

[0034] Furthermore, it also includes a reconstruction loss function:

[0035]

[0036] where Y is the ground truth image, is the enhancement result, λ per is the weight coefficient of the perceptual loss, is to extract multi-layer features using a pre-trained VGG network, and L rec is the reconstruction loss.

[0037] In a second aspect, the present invention provides an infrared image contrast enhancement system for maintaining spatial domain consistency, which is used to implement the above-mentioned infrared image contrast enhancement method for maintaining spatial domain consistency.

[0038] The beneficial effects of the present invention are as follows:

[0039] This method utilizes spatial information interaction and channel information interaction to construct an infrared image enhancement model based on maintaining spatial domain consistency. It extracts spatially consistent expressions from low-resolution features, adjusts the enhancement curve in the feature domain through pixel-by-pixel curves, injects detailed information through local interaction in multi-scale expressions, and trains using a global enhancement loss function, a multi-scale structure loss function, and a histogram loss function. Compared with existing infrared enhancement methods, this method can effectively enhance low-contrast infrared images, effectively improve the spatial domain consistency and detail quality of infrared image enhancement, eliminate problems such as inconsistent pixel intensities in uniform regions of the enhancement result, artifacts, and blurred details in the enhancement result, and improve the visual effect of the enhancement result. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 is a schematic diagram of the overall framework of the present invention.

[0041] Figure 2 is a schematic diagram of the global enhancement module (GEM).

[0042] Figure 3 is a schematic diagram of the local enhancement module (LEM).

[0043] Figure 4 is a schematic diagram of the long-distance spatial information interaction layer and the cross-channel information interaction layer within the global enhancement module of the method of the present invention.

[0044] Figure 5 is a schematic diagram of the detail injector within the local enhancement module of the method of the present invention.

[0045] Figure 6 is a display diagram of the infrared degraded image and the enhancement result of the embodiment of the present invention. (a) Input image, i.e., the infrared degraded image, (b) result of the existing comparison method SNRLLE (published in CVPR2022), (c) result of the method of the present invention, (d) ground truth image. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The present invention will be further described and explained below in conjunction with the specific embodiments. The embodiments are only examples of the present disclosure content and do not delimit the scope of limitation. Without conflict, the technical features of each embodiment of the present invention can be combined accordingly.

[0047] The drawings are only schematic diagrams of the present invention and are not necessarily drawn to scale. Some of the block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0048] The flowchart shown in the accompanying drawings is only an exemplary illustration and does not necessarily include all steps. For example, some steps can be further decomposed, while some steps can be combined or partially combined. Therefore, the actual execution order may be changed according to the actual situation.

[0049] The present invention provides an infrared image contrast enhancement method for maintaining spatial domain consistency, which enhances the contrast of infrared images and improves the image detail quality by using spatial information interaction and channel information interaction. The specific steps are as follows:

[0050] S1: Construct an infrared image enhancement model for maintaining spatial domain consistency, including a feature extraction subnet, a contrast enhancement subnet, and a detail enhancement subnet. Use a global enhancement loss function, a multi-scale structure loss function, and a histogram loss function to train the constructed model until convergence;

[0051] S2: For a given infrared image with low contrast and detail loss (hereinafter referred to as a degraded image), use the infrared image enhancement model described in S1, as Figure 1 shown. The specific steps are as follows:

[0052] S21: Use the feature extraction subnet to encode the image into a low-resolution feature representation;

[0053] S22: Use the global contrast enhancement subnet to enhance the low-resolution feature representation in the feature domain;

[0054] S23: Use the multi-scale detail enhancement subnet to gradually enhance the detail structure and reconstruct the original resolution image.

[0055] The infrared image contrast enhancement model for maintaining spatial domain consistency described in S1 uses a non-linear activation-free network block (NAF BLOCK) as the basic unit to construct the feature extraction subnet described in S1 for extracting low-resolution feature representations. This feature extraction subnet maps the degraded image X into a low-resolution feature representation F L , and inputs it into the contrast enhancement subnet described in S1.

[0056] The contrast enhancement subnet described in S1 is composed of several global enhancement modules (GEMs) in cascade, generating an enhanced feature representation F E . Each GEM maintains the generation of a spatial domain consistency descriptor through long-distance spatial information interaction and cross-channel information interaction, and performs global enhancement with consistency preservation in the feature domain. As Figure 2 shown, the specific calculation process of the GEM is as follows:

[0057] (1) Denote the input of the nth GEM as F n ∈R h×w×c, a two - layer self - attention mechanism is adopted, that is, the spatial interaction layer and the channel interaction layer respectively implement long - distance spatial information interaction and cross - channel information interaction. Among them, the dimension of the attention map of the spatial interaction layer is hw×hw, and the dimension of the attention map of the channel interaction layer is c×c. Then, the outputs of the two layers are added together to obtain the global consistency descriptor P.

[0058] As Figure 4 shown in (a) of n ∈R h×w×c , for the input feature map F s , its dimension is converted into a form suitable for calculating attention, that is, flattened from the feature map dimension h×w×c to hw×c, and then the fully - connected layer f QS is used for spatial interaction encoding. The encoded feature map passes through the linear transformation matrices W KS , W VS , W S respectively to obtain the query matrix Q S , the key matrix K S and the value matrix V s , with the dimension of hw×d s , where d S is the projection dimension; the query matrix Q S , the key matrix K S are used to calculate the attention map with the dimension of hw×hw. The attention map is multiplied by the value matrix V

[0059] As Figure 4 shown in (b) of n ∈R h×w×c , for the input feature map F c , its dimension is converted into a form suitable for calculating attention, that is, flattened from the feature map dimension h×w×c to hw×c, and then the fully - connected layer f QC is used for spatial interaction encoding. The encoded feature map passes through the linear transformation matrices W KC , W VC and is mapped and then deformed to obtain the query matrix Q C , the key matrix K C and the value matrix V C , with the dimension of d c ×hw, where d c is the projection dimension; the query matrix Q C , the key matrix K CCalculate the attention map with dimension d c ×d c and use the attention map and the value matrix V C to multiply the output channel interaction expression, reshape the channel interaction expression back to the dimension of h×w×c, perform residual splicing with the input feature map, and then use the convolutional layer to map the dimension back to h×w×c.

[0060] (2) Use the curve generator to generate an enhanced curve tensor C∈R n from the input F of GEM h ×w×K and the global consistency descriptor P for iterative curve adjustment, where K is the number of curve adjustment iterations.

[0061] In this embodiment, the curve generator adopts a learnable neural network, concatenates the channel dimensions of F n and P as the input of the curve generator, and outputs the enhanced curve tensor C.

[0062] (3) At the k-th curve adjustment, use the k-th channel slice in the enhanced curve tensor C as the enhanced curve parameter, denoted as C k ∈R h×w×1 , and perform the following curve adjustment:

[0063]

[0064] where is the k-th feature domain enhancement result, and ⊙ is the element-wise multiplication;

[0065] The detail enhancement subnet described in S1 is composed of a cascade of several local enhancement modules (LEMs). Each LEM is composed of a resolution reconstructor and a detail injector, which enhance and reconstruct high-resolution details step by step. The enhanced features of the last LEM are decoded into the enhanced result of the original resolution through a convolutional layer. As Figure 3 shown, the specific calculation process of each LEM is as follows:

[0066] (1) Use the resolution reconstructor to preliminarily reconstruct the high-resolution features from the low-resolution enhanced features , denoted as F content . In this embodiment, the dimension of F reconstructed by content is the same as the dimension of , and the resolution reconstructor is an upsampling convolutional layer.

[0067] (2) The detail injector includes a dual attention layer and a weighted fusion operation, and injects the detail structure expression F detail extracted from the original image by the detail extraction layer into F content , and the specific steps are as follows:

[0068] First, in the dual attention layer, F content and F detail serve as the query feature Q, key-value features K and V of neighborhood attention respectively, and calculate the detail perception enhanced feature and the contrast perception enhanced feature Here, F detail is the detail feature extracted from the infrared degraded image by the convolutional layer, which has the same dimension as F content . Multiple convolutional layers for extracting F detail at different scales are called detail extraction layers;

[0069] As Figure 5 shown, the neighborhood attention mechanism multiplies the element Q L at a certain spatial position (x, y) of the query feature Q with the neighborhood elements K and V L within the L window centered at the corresponding spatial position (x, y) of the key-value features K and V to obtain a neighborhood attention map with a spatial size of . Then, multiply the neighborhood attention map with the value feature to obtain the neighborhood perception feature at the position (x, y). Perform the above operations on all positions in the spatial dimension of Q, and the finally obtained neighborhood perception feature is the attention output.

[0070] Secondly, generate the fusion weights of the detail perception enhanced feature and the contrast perception enhanced feature through a weight generator, and then perform weighted fusion on the two to output a high-resolution enhanced feature containing rich detail information.

[0071] In this embodiment, the weight generator adopts a learnable single-layer neural network, and splices and in the channel dimension as the input of the weight generator, and outputs the weight W.

[0072] The loss function during the training of the infrared image enhancement model includes a global enhancement loss function, a local enhancement loss function, and a histogram loss function.

[0073] The reconstruction loss function for the model training described in S1 is defined as where Y and represent the reference ground truth image and the enhanced result of this algorithm respectively, and λ per represent the weight coefficients of the perceptual loss respectively. The perceptual loss is defined as where represents extracting multi-layer features using a pre-trained VGG network.

[0074] The global enhancement loss function L GESupervise the low-resolution enhanced features, and set the low-resolution feature F E with a size of 1 / 2 of the original image M , and the calculation method of the global enhancement loss function is as follows:

[0075]

[0076] Among them, represents the enhancement result generated from the low-resolution feature, represents a convolutional decoding layer, and Y ↓M represents downsampling the ground-truth image to 1 / 2 of the original image size M times to match the size of the low-resolution enhancement result.

[0077] The local enhancement loss function L used for the model training described in S1 LE performs multi-scale supervision on the features of, and sets the multi-scale feature with a size of 1 / 2 of the original image M-m , and the calculation method of the local enhancement loss function is as follows:

[0078]

[0079] Among them, represents the enhancement result generated from the multi-scale feature, represents a convolutional decoding layer, and Y ↓M represents downsampling the ground-truth image to 1 / 2 of the original size m times to match the size of the multi-scale enhancement result.

[0080] The histogram loss function L used for the model training described in S1 hist performs histogram supervision on the enhancement result. The present invention uses a soft histogram operator to calculate the image histogram, and the calculation method of the histogram loss function is as follows:

[0081]

[0082] Among them, h represents the histogram of the model enhancement result, represents the histogram of the ground-truth image.

[0083] Figure 6 is the operation result of a specific embodiment based on the method of the present invention. Compared with the existing cutting-edge infrared enhancement method SNRLLE published in CVPR2022, this method can effectively enhance low-contrast infrared images, effectively improve the spatial consistency and detail quality of infrared image enhancement, eliminate the problems of inconsistent pixel intensity in the uniform area of the enhancement result, artifacts, and detail blurring, and effectively improve the visual effect of the enhancement result.

[0084] In this embodiment, an infrared image contrast enhancement system for maintaining spatial domain consistency is further provided, including:

[0085] An infrared image enhancement model module, which includes a feature extraction subnet, a contrast enhancement subnet, and a detail enhancement subnet. The feature extraction subnet is used to encode a given infrared degraded image with low contrast and detail loss into a low-resolution feature representation. The contrast enhancement subnet is used to enhance the low-resolution feature representation in the feature domain to obtain an enhanced feature representation. The detail enhancement subnet is used to gradually enhance the detail structure of the enhanced feature representation and reconstruct the original resolution image.

[0086] A model training module, which is used to train an infrared image enhancement model.

[0087] A contrast enhancement module, which is used to enhance the contrast of a given infrared degraded image by using the trained infrared image enhancement model and output an enhancement result.

[0088] For the system embodiment, since it basically corresponds to the method embodiment, the relevant parts can refer to the partial description of the method embodiment, and the implementation methods of the remaining modules will not be elaborated here. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0089] The embodiments of the system of the present invention can be applied to any device with data processing capabilities, and the any device with data processing capabilities can be a device or apparatus such as a computer. The system embodiment 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 the processor of any device with data processing capabilities reading the corresponding computer program instructions in the non-volatile memory into the memory for running.

[0090] The above embodiments only represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. For those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. An infrared image contrast enhancement method for maintaining airspace consistency, characterized in that, Including the following steps: Training an infrared image enhancement model, where the infrared image enhancement model includes a feature extraction subnet, a contrast enhancement subnet, and a detail enhancement subnet; The feature extraction subnet is used to encode a given infrared degraded image with low contrast and detail loss into a low-resolution feature representation, and the contrast enhancement subnet is used to enhance the low-resolution feature representation in the feature domain to obtain an enhanced feature representation; The detail enhancement subnet is used to gradually enhance the detail structure of the enhanced feature representation and reconstruct the original-resolution image; Using the trained infrared image enhancement model to enhance the contrast of a given infrared degraded image and output an enhanced result.

2. The method for enhancing the contrast of an infrared image while maintaining airspace consistency according to claim 1, wherein The feature extraction subnet uses a non-linear activation-free network block.

3. The infrared image contrast enhancement method for maintaining airspace consistency according to claim 1, characterized in that The contrast enhancement subnet is composed of a cascade of several global enhancement modules (GEMs). Each GEM maintains the generation of a spatial consistency descriptor through long-distance spatial information interaction and cross-channel information interaction, and performs global enhancement with consistency preservation in the feature domain.

4. The infrared image contrast enhancement method for maintaining airspace consistency according to claim 3, characterized in that, The calculation process of the GEM is as follows: Denote the input of the nth GEM as F n ∈R h×w×c , where h, w, and c are the length, width, and number of channels of the input image respectively; a two-layer self-attention mechanism is adopted, that is, the spatial interaction layer and the channel interaction layer are used to achieve long-range spatial information interaction and cross-channel information interaction respectively. The dimension of the attention map of the spatial interaction layer is hw×hw, and the dimension of the attention map of the channel interaction layer is c×c. Then, the outputs of the two layers are added together to obtain the global consistency descriptor P; Generate an enhanced curve tensor C ∈ R from the input F of GEM using a curve generator n and the global consistency descriptor P, where K is the number of curve adjustment iterations; h×w×K ​ Use the enhanced curve tensor C for F n Perform iterative curve adjustment, and use the result of the last adjustment as the output.

5. The method for enhancing the contrast of an infrared image while maintaining airspace consistency according to claim 4, wherein During iterative curve adjustment, the k-th curve adjustment uses the k-th channel slice in the enhanced curve tensor C as the enhanced curve parameter, denoted as C k ∈R h×w×1 , and perform the following curve adjustment: Among them, is the k-th feature domain enhancement result, and ⊙ is element-wise multiplication; 6. The method for enhancing the contrast of an infrared image for maintaining airspace consistency according to claim 1, wherein The detail enhancement subnet is composed of a cascade of several local enhancement modules (LEMs). Each LEM consists of a resolution reconstructor and a detail injector to gradually enhance and reconstruct high-resolution details; the enhanced feature of the last LEM is decoded into an enhanced result of the original resolution through a convolutional layer.

7. The method for enhancing the contrast of an infrared image while maintaining airspace consistency according to claim 6, characterized in that The calculation process of the LEM is as follows: Reconstruct high-resolution features from low-resolution enhanced features using a resolution reconstructor Preliminarily reconstruct high-resolution features, denoted as F content , from The F reconstructed content has the same dimension as ; the resolution reconstructor is an upsampling convolutional layer are the input and output of the m-th LEM respectively The detail injector includes a dual attention layer and a weighted fusion operation. First, in the dual attention layer, F content and F detail are the query feature Q, key-value features K and V of neighborhood attention respectively, and the detail-aware enhanced feature and the contrast-aware enhanced feature are calculated, where F detail is the detail feature with the same dimension as F content extracted from the infrared degraded image by the convolutional layer; Second, the fusion weights of the detail-aware enhanced feature and the contrast-aware enhanced feature are generated by the weight generator, and then the two are weighted and fused to output the high-resolution enhanced feature 8. The method for enhancing the contrast of an infrared image while maintaining airspace consistency according to claim 7, characterized in that The loss function during the training of the infrared image enhancement model includes a global enhancement loss function, a local enhancement loss function, and a histogram loss function; Global enhanced loss function L GE is as follows: Among them, represents the enhanced result generated from the low-resolution feature representation, represents a convolutional decoding layer, Y ↓M represents downsampling the ground-truth image to 1 / 2 M times the original image size to match the size of the low-resolution enhanced result; ‖.‖1 is the L1 norm; Set multi-scale features The size is 1 / 2 of the original image M-m , the local enhancement loss function L LE is as follows: where M is the number of LEMs, represents the enhanced result generated from multi-scale features, Y ↓M represents downsampling the ground truth image to 1 / 2 of the original size m times to match the size of the multi-scale enhancement result, and SSIM represents structural similarity; Histogram loss function L hist is as follows: where h represents the histogram of the enhancement result, represents the histogram of the ground truth image.

9. The infrared image contrast enhancement method for maintaining airspace consistency according to claim 1, characterized in that It also includes a reconstruction loss function: Among them, Y is the true value image, is the enhancement result, λ per is the weight coefficient of the perceptual loss, is to extract multi-layer features using the pre-trained VGG network, L rec is the reconstruction loss.

10. An infrared image contrast enhancement system for maintaining airspace consistency, which is used to implement the method described in claim 1, characterized in that, The system includes: An infrared image enhancement model module, which includes a feature extraction subnet, a contrast enhancement subnet, and a detail enhancement subnet. The feature extraction subnet is used to encode a given infrared degraded image with low contrast and detail loss into a low-resolution feature representation, and the contrast enhancement subnet is used to enhance the low-resolution feature representation in the feature domain to obtain an enhanced feature representation; the detail enhancement subnet is used to gradually enhance the detail structure of the enhanced feature representation and reconstruct the original-resolution image; A model training module, which is used to train an infrared image enhancement model; A contrast enhancement module, which is used to use the trained infrared image enhancement model to enhance the contrast of a given infrared degraded image and output an enhanced result.

Images