Method for multi-scale infrared polarization image enhancement based on MSD-PCNN

Abstract

The application discloses a multi-scale infrared polarization image enhancement method based on MSD-PCNN and relates to the technical field of infrared image processing. First, based on non-subsampled shearlet transform (NSST), an input intensity image I and a polarization feature map DoLP are decomposed in a multi-scale and multi-directional manner to obtain low-frequency subbands and high-frequency subbands. Second, a low-frequency subband fusion rule based on distance difference perception is designed according to the obtained low-frequency subbands. Finally, a high-frequency subband fusion rule based on MSD-PCNN is designed according to the obtained high-frequency subbands to fully extract texture and edge detail features. The multi-scale infrared polarization image enhancement method based on MSD-PCNN provided by the application suppresses redundant information interference, improves the distinctness of image structure levels, and also improves the visual comfort and global naturalness of a fused image.

Description

Technical Field

[0001] This invention relates to the field of infrared image processing technology, and in particular to a technology based on...

[0002] MSD-PCNN is a multi-scale infrared polarization image enhancement method. Background Technology

[0003] Infrared polarization image fusion enhancement significantly improves target detection efficiency by matching complementary composite image features and leveraging information synergy. This method not only overcomes the limitations of single-modal imaging in low-contrast, weak-texture scenes but also makes the reconstructed image more consistent with human visual perception mechanisms, providing a solid foundation for subsequent image recognition, target extraction, and other processing tasks. In military-civilian integration fields such as reconnaissance and early warning, maritime search and rescue, and disaster prevention and control, this technology demonstrates irreplaceable strategic application value.

[0004] From an information-level perspective, multi-scale transform (MST) methods, as an important category of pixel-level processing, have attracted widespread attention and become one of the most representative research directions in this field due to their low computational complexity and ability to preserve source image information well. Typical MST techniques include Laplacian Pyramid (LP), Wavelet Transform (WT), Dual-Tree Complex Wavelet Transform (DTCWT), and geometric multi-scale analysis methods such as Nonsubsampled Contourlet Transform (NSCT) and Nonsubsampled Shearlet Transform (NSST). These methods have all been widely applied in numerous studies.

[0005] In comparative experiments conducted by our research team in 2024, we found that the traditional NSCT method exhibits significant limitations when handling complex textures. For example, when the radius of curvature of the target edge is less than 3 pixels, the peak signal-to-noise ratio (PSNR) of the reconstructed image decreases by approximately 6 dB. This phenomenon prompted us to shift our research focus to a new generation of improved methods. In earlier research, S. Wang et al. creatively combined NSCT with convolutional neural networks (CNNs). However, in practical applications, it was found that this single-filter architecture leads to the loss of texture information and causes significant brightness diffusion when reconstructing high-frequency details. In 2014, Kong's team proposed a fusion framework combining NSST decomposition and pulse-coupled neural networks (PCNNs) and introduced a spatial frequency modulation mechanism. This method demonstrated excellent performance in infrared and visible light image fusion tasks and opened up new research avenues for the modulation of PCNN metrics. In 2016, Z. Zhou's team proposed a Hybrid Multi-Scale Decomposition (HMSD) method, employing a dual-filter parallel architecture: bilateral filters preserve fine textures, while Gaussian filters focus on extracting large-scale edges. In 2018, Z. Jin et al. proposed a gradient-domain guided filtering scheme, effectively improving visual perception quality, but still exhibiting artifact problems in non-uniform lighting scenes. In 2022, G. Liu et al. introduced Gaussian curvature filtering into the HMSD framework and, through a curvature-driven adaptive mechanism, achieved edge sharpness without parametric modeling. In 2023, D. Zou's team combined guided image filtering (GIF) with side-window guided filtering (SGIF), enhancing not only edge sharpness but also improving the extraction of weak texture features based on the HMSD architecture.

[0006] While existing infrared polarization image fusion algorithms have made some progress in target enhancement and structure discrimination, they still have many shortcomings. Most existing infrared polarization image fusion algorithms rely on a multi-scale transform (MST) framework with a fixed kernel function. Essentially, this model is based on linear operations to construct feature representation models, making it difficult to adapt to real-world images with highly nonlinear characteristics. When faced with complex structures and multimodal information, this method is prone to feature mismatch and information mixing, significantly weakening the discriminative ability of the fused image. As the scale increases progressively, polarization features gradually decay in the multi-scale domain, while radiance information continues to dominate, resulting in a decrease in the ability to extract low-intensity and weak-texture features. Simultaneously, linear fusion mechanisms have limited performance in distinguishing effective targets from background clutter, often exhibiting block artifacts and noise amplification problems, further interfering with the accurate representation of salient regions and even obscuring key polarization features. Furthermore, traditional MST methods lack stability in different task scenarios, exhibiting strong dependence on specific imaging conditions and making it difficult to maintain consistent fusion quality in variable environments.

[0007] Overall, the current MST-dominated fusion strategy suffers from core bottlenecks such as weak robustness, poor adaptability, and limited information discrimination ability in highly nonlinear and complex structural contexts. Summary of the Invention

[0008] The purpose of this invention is to provide a multi-scale infrared polarization image enhancement method based on MSD-PCNN, which solves the problems of weak robustness, poor adaptability and limited information discrimination ability of traditional fusion methods in the context of high nonlinearity and complex structure.

[0009] To achieve the above objectives, this invention provides a multi-scale infrared polarization image enhancement method based on MSD-PCNN, comprising the following steps:

[0010] Step 1: Based on the non-subsampled shear wave transform (NSST), perform multi-scale and multi-directional decomposition on the input intensity image I and polarization feature map DoLP to obtain the low-frequency subband and high-frequency subband;

[0011] Step 2: Based on the low-frequency sub-bands obtained in Step 1, design low-frequency sub-band fusion rules based on distance difference perception;

[0012] Step 3: Based on the high-frequency subbands obtained in Step 1, design a high-frequency subband fusion rule based on MSD-PCNN to fully extract texture and edge detail features.

[0013] Preferably, in step 1, the original image is divided into a set of low-frequency sub-bands L by NSST decomposition. DoLP L I and several high-frequency subbands Where s represents scale and k represents direction; the low-frequency subband contains the overall brightness and structural information of the image, while the high-frequency subband focuses on edge, detail and texture features.

[0014] Preferably, the process of designing low-frequency sub-band fusion rules based on distance difference perception in step 2 is as follows:

[0015] S21. Calculate the low-frequency subband L corresponding to the input intensity image I and polarization feature map DoLP, respectively. I and L DoLP Statistical characteristics and the corresponding pixel mean Mean(L I Mean(L) DoLP ) and Median(L I Median(L) DoLP The sum of these values, used as a global feature reference value, is expressed as follows:

[0016]

[0017]

[0018] S22. Construct a response-sensitive factor. Based on the difference between each pixel value and the global feature value, calculate the saliency weight through an exponential function to enhance the response of the structural region.

[0019] S23. Perform normalized weighted fusion; use the response sensitivity factor as the weight to perform pixel-level fusion on the two low-frequency sub-bands to obtain the reconstructed low-frequency image.

[0020] Preferably, the expression for calculating the significance weight using the exponential function in S22 is as follows:

[0021]

[0022] In the formula, These represent subband L based on DoLP. DoLP The weighting factor and low-frequency subband L calculated at pixel position (i,j) I The weighting factor is calculated at pixel position (i,j), where m represents the adjustment parameter in the exponential function, and L... DoLP (i,j) represents the grayscale feature value of the DoLP low-frequency component at pixel (i,j), L I (i,j) represents the gray-level feature value of the low-frequency component of the intensity image at pixel (i,j).

[0023] Preferably, the reconstructed low-frequency image L obtained in S23 recon The expression for (i,j) is as follows:

[0024]

[0025] In the formula, L DoLP (i,j) represents the grayscale feature value of the DoLP low-frequency component at pixel (i,j), L I (i,j) represents the gray-level feature value of the low-frequency component of the intensity image at pixel (i,j).

[0026] Preferably, the process of designing the high-frequency subband fusion rule based on MSD-PCNN in step 3 is as follows:

[0027] S31. Perform multi-scale, multi-directional high-frequency subband decomposition on the DoLP and I images to obtain the coefficients in the k-th direction at the s-th scale. Then, multi-scale morphological difference (MSD) is used to enhance the structural features of each high-frequency subband; based on morphological structural elements (SE) at different scales, multi-scale gradient feature maps are extracted by combining dilation and erosion operations. The expression is as follows:

[0028]

[0029] In the formula, t is the maximum scale layer number, SE s For the morphological structuring elements at scale s, Dilate represents the dilation operation, used to expand bright areas in the image, and Erode represents the erosion operation, used to shrink bright areas in the image.

[0030] S32. Using the MSD feature map as the input to the Pulse Coupled Neural Network (PCNN), an iterative neuron model is constructed for each pixel based on the chosen parameter set Para, and the corresponding pulse firing time map is output, as shown in the following expression:

[0031]

[0032] Among them, Y n (i,j) represents the binary activation state of pixel (i,j) at the nth iteration. The cumulative pulse count characterizes the importance of the pixel region, and the larger the cumulative time map, the more significant the region is.

[0033] S33. Construct the decision mapping; based on the pulse time intensity at the same position in the two input images, construct the decision mapping M. s,k (i,j);

[0034] S34, Decision Fusion: Perform pixel-level discriminative fusion based on the decision mapping to generate the final high-frequency fused subband H. s,k (i,j);

[0035] S35. Iterative loop: For all scales s = 1, 2, 3, 4, 5 and the corresponding directional subbands k = 1, 2, ..., K(s), iterate through the above process to obtain the complete fused high-frequency coefficient set H. recon .

[0036] Preferably, the expression for the decision mapping constructed in S33 is as follows:

[0037]

[0038] In the formula, This represents the DoLP pulse temporal intensity at position (i,j), scale s, and source image k. This represents the total pulse temporal intensity of the source image at position (i,j), at the s-th scale, and the k-th source image.

[0039] Preferably, the final high-frequency fusion subband H is generated in S34. s,k The expression for (i,j) is as follows:

[0040]

[0041] In the formula, This represents the DoLP high-frequency subband of the source image at position (i,j), at the s-th scale. This represents the high-frequency subband of the I-map at position (i,j), scale s, and source image k.

[0042] Preferably, the complete fusion high-frequency coefficient set H in S35 recon The expression is as follows:

[0043]

[0044] In the formula, K(s) represents the directional sub-bands corresponding to each scale, k = 1, 2, ..., K(s).

[0045] Therefore, the multi-scale infrared polarization image enhancement method based on MSD-PCNN described above has the following beneficial effects:

[0046] (1) Multi-source polarization feature synergistic enhancement improves the ability to identify target saliency; This invention introduces two types of polarization feature maps, DoLP and I, and achieves feature synergistic expression in the multi-scale shear wave domain through complementary modeling of intensity features and polarization response, effectively improving the structural fidelity and polarization contrast of the target area, and significantly enhancing the saliency and detectability of weak targets in complex backgrounds.

[0047] (2) Based on the high-frequency decision mechanism of MSD-PCNN, the edge and texture expression are enhanced. This invention adopts a high-frequency fusion strategy that combines multi-scale morphological gradient extraction with pulse-coupled neural network. Through TimeMap temporal mapping, the structural features are dynamically activated and the decision selection is realized, which fully preserves the continuity of detail edges and the clarity of directional texture, suppresses redundant information interference, and improves the image structure hierarchy.

[0048] (3) Low-frequency region difference perception weighted fusion to ensure global brightness consistency; This invention designs a low-frequency weighted mechanism driven by statistical feature difference, uses exponential response factor to perceive regional saliency, realizes pixel-level adaptive fusion, effectively suppresses artifacts in flat regions, maintains image brightness consistency, and improves the visual comfort and global naturalness of the fused image.

[0049] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0050] Figure 1 This is a diagram illustrating the overall architecture of the multi-scale infrared polarization image enhancement method based on MSD-PCNN of this invention.

[0051] Figure 2 This is a diagram illustrating the joint reconstruction framework of NSST decomposition and PCNN temporal mapping according to an embodiment of the present invention.

[0052] Figure 3 This is a 0° takeoff diagram of a mountain drone according to an embodiment of the present invention;

[0053] Figure 4 This is a 60° takeoff diagram of a mountain drone according to an embodiment of the present invention;

[0054] Figure 5 This is a 120° takeoff diagram of a mountain drone according to an embodiment of the present invention;

[0055] Figure 6 This is a DoLP diagram of a mountain drone according to an embodiment of the present invention;

[0056] Figure 7 Figure I shows a mountain drone according to an embodiment of the present invention;

[0057] Figure 8 This is an enhanced reconstruction R diagram according to an embodiment of the present invention. Detailed Implementation

[0058] The following detailed description of embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0059] Please see Figures 1-8 A multi-scale infrared polarization image enhancement method based on MSD-PCNN is proposed. The overall framework includes an infrared polarization image input module, an NSST decomposition and morphological gradient processing module, and a reconstruction and enhancement output module. The specific process is as follows:

[0060] Step 1: Input the intensity image I and the polarization feature map DoLP into... Figure 2 In the framework shown, multi-scale and multi-directional decomposition is first performed using Non-Subsampled Shear Wavelet Transform (NSST). NSST, as an improved wavelet transform method, possesses translation invariance and multi-directionality, effectively preserving edge and texture details of the image while avoiding information loss caused by downsampling in traditional multi-scale transforms. Through NSST decomposition, the original image is divided into a set of low-frequency subbands L. DoLP L I and several high-frequency subbands Where 's' represents the scale and 'k' represents the direction. The low-frequency sub-band mainly contains the overall brightness and structural information of the image, while the high-frequency sub-band focuses on edge, detail, and texture features. This decomposition method not only facilitates subsequent feature extraction and weight calculation but also provides flexibility for the design of fusion strategies, enabling differentiated processing methods for different frequency bands, thereby achieving a balance between brightness consistency and detail preservation during the reconstruction stage. By performing non-subsampled shear wave transform (NSST) on the two input feature images, a superior multi-scale geometric decomposition is achieved, laying a solid feature foundation for multi-source information fusion and providing accurate input representations for subsequent specialized processing of high-frequency details and low-frequency structures, as well as adaptive weighting strategies.

[0061] Step 2: Based on the low-frequency sub-bands obtained in Step 1, design low-frequency sub-band fusion rules based on distance difference perception; the specific process is as follows:

[0062] S21. Calculate the low-frequency subband L corresponding to the input intensity image I and polarization feature map DoLP, respectively. I and L DoLP Statistical characteristics and the corresponding pixel mean Mean(L I Mean(L) DoLP ) and Median(L I Median(L) DoLP The sum of these values, used as a global feature reference value, is expressed as follows:

[0063]

[0064]

[0065] S22. Construct a response sensitivity factor. Based on the difference between each pixel value and the global feature value, calculate the saliency weight using an exponential function to enhance the response of the structural region; the expression is as follows:

[0066]

[0067] In the formula, These represent subband L based on DoLP. DoLP The weighting factor and low-frequency subband L calculated at pixel position (i,j) I The weighting factor is calculated at pixel position (i,j), where m represents the adjustment parameter in the exponential function, and L... DoLP (i,j) represents the grayscale feature value of the DoLP low-frequency component at pixel (i,j), L I (ij) represents the gray-level feature value of the low-frequency component of the intensity image at pixel (i,j).

[0068] S23. Perform normalized weighted fusion; use the response sensitivity factor as weight to perform pixel-level fusion on the two low-frequency sub-bands to obtain the reconstructed low-frequency image; wherein, the obtained reconstructed low-frequency image L recon The expression for (i,j) is as follows:

[0069]

[0070] In the formula, L DoLP (i,j) represents the grayscale feature value of the DoLP low-frequency component at pixel (i,j), L I (i,j) represents the grayscale feature value of the low-frequency component of the intensity image at pixel (i,j). Through the calculation of the formula, the response to structurally prominent regions can be effectively improved, while interference from flat regions can be suppressed, achieving more robust low-frequency fusion, thus obtaining the final fused low-frequency image output L. recon ;

[0071] Step 3: Based on the high-frequency subbands obtained in Step 1, design high-frequency subband fusion rules based on MSD-PCNN to fully extract texture and edge detail features; the specific process is as follows:

[0072] S31. Perform multi-scale, multi-directional high-frequency subband decomposition on the DoLP and I images to obtain the coefficients in the k-th direction at the s-th scale. Then, multi-scale morphological difference (MSD) is used to enhance the structural features of each high-frequency subband; based on morphological structural elements (SE) at different scales, multi-scale gradient feature maps are extracted by combining dilation and erosion operations. The expression is as follows:

[0073]

[0074] In the formula, t is the maximum scale layer number, SE s For the morphological structuring elements at scale s, Dilate represents the dilation operation, used to expand bright areas in the image, and Erode represents the erosion operation, used to shrink bright areas in the image.

[0075] S32. Using the multi-scale morphological gradient (MSD) feature map as the input to the pulse-coupled neural network (PCNN), an iterative neuron model is constructed for each pixel based on the chosen parameter set Para, and the corresponding pulse firing time map (TimeMap) is output, as shown in the following expression:

[0076]

[0077] Among them, Y n(i,j) represents the binary activation state of pixel (i,j) at the nth iteration. The cumulative pulse count characterizes the importance of this pixel region; a larger cumulative timescale indicates a more significant region. A decision rule is established using PCNN pulse behavior features to preferentially preserve salient regions.

[0078] S33. Construct the decision mapping; based on the pulse time intensity at the same position in the two input images, construct the decision mapping M. s,k (i,j); The specific expression is as follows:

[0079]

[0080] In the formula, This represents the DoLP pulse temporal intensity at position (i,j), scale s, and source image k. This represents the total pulse temporal intensity of the source image at position (i,j), at the s-th scale, and the k-th source image.

[0081] S34, Decision Fusion: Pixel-level discriminative fusion is performed based on the decision mapping. This pixel-level fusion method effectively maintains the continuity of edge and texture details, generating the final high-frequency fusion sub-band H. s,k (i,j); The specific expression is as follows:

[0082]

[0083] In the formula, This represents the DoLP high-frequency subband of the source image at position (i,j), at the s-th scale. This represents the high-frequency subband of the I-map at position (i,j), scale s, and source image k.

[0084] S35. Iterative loop: For all scales s = 1, 2, 3, 4, 5 and the corresponding directional subbands k = 1, 2, ..., K(s), iterate through the above process to obtain the complete fused high-frequency coefficient set H. recon The specific expression is as follows:

[0085]

[0086] In the formula, K(s) represents the corresponding direction k = 1, 2, ..., K(s) at each scale.

[0087] This invention takes a mountain drone deployment scenario as the experimental target. First, it acquires data using a long-wave infrared polarization camera. Figures 3 to 5 Three characteristic maps showing the initial deviation at 0°, 60°, and 120° were generated. Then, the physical quantity DoLP obtained from these maps was calculated. Figure 6 ) and I( Figure 7The algorithm uses this as input to address the issues of incomplete information and limited target representation in DoLP. Furthermore, to address the insufficient adaptability of the nonlinear structure in MST due to its fixed transform kernel, a statistically constrained reconstruction strategy is introduced in the low-frequency region, while a biomimetic visual attention mechanism is incorporated in the high-frequency region. This guides MSD-PCNN to adaptively extract directional feature subbands with symmetry and compactness, thereby achieving high-precision reconstruction of detailed structures and obtaining the final enhanced reconstructed image R( Figure 8 ).

[0088] Therefore, this invention employs the aforementioned multi-scale infrared polarization image enhancement method based on MSD-PCNN, which suppresses redundant information interference, improves the clarity of image structure hierarchy, and also enhances the visual comfort and global naturalness of the fused image.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-scale infrared polarization image enhancement method based on MSD-PCNN, characterized in that, Includes the following steps: Step 1: Based on the non-subsampled shear wave transform (NSST), perform multi-scale and multi-directional decomposition on the input intensity image I and polarization feature map DoLP to obtain the low-frequency subband and high-frequency subband; Step 2: Based on the low-frequency sub-bands obtained in Step 1, design low-frequency sub-band fusion rules based on distance difference perception; Step 3: Based on the high-frequency subbands obtained in Step 1, design high-frequency subband fusion rules based on MSD-PCNN to fully extract texture and edge detail features; In step 1, the original image is divided into a set of low-frequency subbands through NSST decomposition. , and several high-frequency subbands , where s represents the scale and k represents the direction; The low-frequency subband contains overall brightness and structural information of the image, while the high-frequency subband focuses on edge, detail, and texture features. The process of designing the low-frequency subband fusion rule based on distance difference awareness in step 2 is as follows: S21. Calculate the low-frequency subbands corresponding to the input intensity image I and polarization feature map DoLP, respectively. and Statistical characteristics 、 and the corresponding pixel average 、 With median 、 The sum, used as a global feature reference value, is expressed as follows: (1) (2)； S22. Construct a response sensitivity factor and calculate the significance weights using an exponential function based on the differences between each pixel value and the global feature value. S23. Perform normalized weighted fusion; use the response sensitivity factor as the weight to perform pixel-level fusion on the two low-frequency sub-bands to obtain the reconstructed low-frequency image.

2. The multi-scale infrared polarization image enhancement method based on MSD-PCNN according to claim 1, characterized in that, The expression for calculating significance weights using the exponential function in S22 is as follows: (3) (4) In the formula, Indicates DoLP-based subband At pixel position The calculated weighting factor is used to measure the significance of this position in low-frequency fusion. Indicates low-frequency subband At pixel position The calculated weighting factors This represents the adjustment parameter in the exponential function, used to control the sensitivity of the weights to pixel differences. Indicates the low-frequency components of DoLP in pixels grayscale feature value at that location This indicates the low-frequency components of the intensity image at the pixel level. The grayscale feature value at that location.

3. The multi-scale infrared polarization image enhancement method based on MSD-PCNN according to claim 2, characterized in that: The reconstructed low-frequency image obtained in S23 The expression is as follows: (5) In the formula, Indicates the low-frequency components of DoLP in pixels grayscale feature value at that location Indicates the low-frequency components of the intensity image at the pixel level. The grayscale feature value at that location.

4. The multi-scale infrared polarization image enhancement method based on MSD-PCNN according to claim 3, characterized in that: The process of designing the high-frequency subband fusion rule based on MSD-PCNN in step 3 is as follows: S31. Perform multi-scale, multi-directional high-frequency subband decomposition on the DoLP and I images to obtain the coefficients in the k-th direction at the s-th scale. Then, multi-scale morphological difference (MSD) is used to enhance the structural features of each high-frequency subband. Based on morphological structural elements (SE) at different scales, multi-scale gradient feature maps are extracted by combining dilation and erosion operations. The expression is as follows: (6) In the formula, t is the maximum scale layer number. For morphological structural elements at scale s, This represents the dilation operation, used to expand bright areas in an image. This indicates an erosion operation, used to reduce bright areas in an image; S32. Using the MSD feature map as the input to the Pulse Coupled Neural Network (PCNN), an iterative neuron model is constructed for each pixel based on the chosen parameter set Para, and the corresponding pulse firing time map is output, as shown in the following expression: (7) in, For the pixel at the nth iteration The binary activation state, the cumulative number of pulses characterizes the importance of the pixel region, and the larger the cumulative time map, the more significant the region is; S33. Construct a decision map; based on the pulse time intensity at the same position in the two input images, construct a decision map. ; S34, Decision Fusion: Perform pixel-level discriminative fusion based on the decision mapping to generate the final high-frequency fused subband. ; S35. Iterative loop: For all scales s=1,2,3,4,5 and the corresponding directional subbands k=1,2,…,K(s) at each scale, iterate through the above process to obtain a complete set of fused high-frequency coefficients. .

5. The multi-scale infrared polarization image enhancement method based on MSD-PCNN according to claim 4, characterized in that, The expression for the decision mapping constructed in S33 is as follows: (8) In the formula, Indicates the location Above, the first The first scale, the first DoLP pulse temporal intensity of each source image Indicates the location Above, the first The first scale, the first Total pulse time intensity of each source image.

6. The multi-scale infrared polarization image enhancement method based on MSD-PCNN according to claim 5, characterized in that, The final high-frequency fusion subband is generated in S34. The expression is as follows: (9) In the formula, Indicates the location Above, the first The first scale, the first DoLP high-frequency subbands of each source image Indicates the location Above, the first The first scale, the first The high-frequency subband of the I-image of a source image.

7. The multi-scale infrared polarization image enhancement method based on MSD-PCNN according to claim 6, characterized in that, Complete fusion high-frequency coefficient set in S35 The expression is as follows: (10) In the formula, Let k = 1, 2, ..., K(s) represent the directional subbands at each scale.

Images