Infrared dim small target detection method based on target feature multi-perception

Abstract

The application discloses an infrared dim small target detection method based on target feature multi-perception, and relates to the technical field of infrared image target detection. The method comprises the following steps: two new target region feature images are respectively constructed by using morphological erosion and dilation operations; a DoG filter is used to filter the two new target region feature images, and the filtering results are fused to obtain a fused target region feature image; a structure tensor is used to extract corner features from the fused target region feature image to obtain a target response map; and false targets in the target response map are removed by using a target candidate region image to obtain a target detection result. The application fuses and improves key steps of multiple single infrared images, and overcomes defects of a single detection algorithm.

Description

Technical Field

[0001] This invention relates to the field of infrared image target detection technology, and more specifically to a method for detecting weak infrared targets using multiple sensing of target features. Background Technology

[0002] Infrared weak target detection is an image processing technique that accurately detects infrared weak targets with a size no larger than 9×9 pixels from an infrared background. It is currently widely used in various military and civilian fields, such as early warning systems, precision-strike weapons, and forest fire prevention. Compared with visible light and other radar bands, the infrared band has advantages such as passive identification, good concealment performance, high resolution, long detection range, and all-weather operation. Therefore, infrared images have become one of the research hotspots in the field of image processing. Due to the long imaging distance and the influence of background clutter, small targets in infrared images usually appear as weak bright spots at the pixel level, lacking texture and shape information. In complex backgrounds, they often exhibit a low signal-to-clutter ratio, with the target almost completely submerged in background clutter. Utilizing the only prior features of the background or target to achieve effective detection / segmentation of infrared weak targets is of great significance. Research on single-frame detection algorithms for infrared weak targets is relatively widespread, and this convenient and efficient detection method is more in line with practical application needs.

[0003] Current single-frame detection methods can be divided into four types: methods based on background consistency features, methods based on target saliency features, methods based on low-rank and sparse decomposition of image structure, and methods based on deep learning. Methods based on background consistency features typically utilize relevant filtering operations to obtain an image containing only the background scene. The image containing only the target is obtained by subtracting the original image from the background image. This algorithm is only suitable for target detection against simple, flat backgrounds, and performs poorly in complex backgrounds with a high false alarm rate. Methods based on target saliency are more diverse, mainly using the saliency features of the target to distinguish it from the background, enhancing target contrast while suppressing the influence of background clutter. However, different target saliency extraction strategies in actual algorithms may lead to reduced computational efficiency and algorithm robustness. Methods based on low-rank and sparse decomposition of image structure transform the target detection problem into an optimization problem of low-rank and sparse matrix decomposition. They often consume a lot of time in tensor construction, image reconstruction, and optimization iteration, resulting in reduced computational efficiency. Furthermore, the detection performance is unstable in low signal-to-clutter and high-brightness noisy scenes. Deep learning-based methods require a large amount of data support. Currently, due to the sensitivity of data, it is difficult to obtain a large amount of training sample data. Moreover, because the proportion of targets is extremely low, deep learning network models are prone to failing to learn the features of weak infrared targets and instead learning background information, leading to reduced detection performance.

[0004] Therefore, how to provide a new infrared weak target detection algorithm that enhances background suppression and target enhancement effects, and improves computational efficiency is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of this, the present invention provides an infrared weak target detection method based on multiple sensing of target features, which belongs to the single-frame detection method based on target saliency features. It integrates and improves the key operations of the method based on background consistency features, the method based on target saliency features, and the method based on low-rank and sparse decomposition of image structure to solve the technical problems existing in the background technology.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A method for detecting weak infrared targets using multiple sensing of target features, comprising the following steps:

[0008] Step 1: Construct two new target region feature images using morphological erosion and dilation operations respectively;

[0009] Step 2: Use the DoG filter to filter the two new target region feature images, and fuse the filtering results to obtain the fused target region feature image;

[0010] Step 3: Use the structure tensor to extract corner features from the fused target region feature image to obtain the target response map;

[0011] Step 4: Use the target candidate region image to remove false targets in the target response map to obtain the target detection result.

[0012] Preferably, in step one, two new target region feature images are constructed using morphological erosion and dilation operations, specifically including:

[0013] S11. Perform morphological erosion and dilation operations on the input infrared image I to obtain the eroded image I. e and dilated image I d

[0014] I e =I⊙SE

[0015]

[0016] Among them, “⊙” and These represent the symbols for corrosion and expansion operations, respectively, and SE represents a structural element with a radius of 1.

[0017] S12, Using the input infrared image I and the erosion image I e The difference value is used to obtain the compactness mapping diagram I. c

[0018] I c =II e

[0019] Using dilated image I d The heterogeneity mapping I is obtained by the difference between the input infrared image I and the input infrared image I. h

[0020] I h =I d -I

[0021] The compact mapping diagram I c Heterogeneity Mapping I h These are two new target region feature images.

[0022] Preferably, step two specifically includes:

[0023] S21. Based on the DoG filter, obtain two DoG cores with different standard deviations, DoG1 core and DoG2 core. The σ1 and σ2 in the DoG1 core are set to 2 and 0.5 respectively, and the σ1 and σ2 in the DoG2 core are set to 2 and 3 respectively.

[0024] S22. Using the DoG1 kernel to map the heterogeneity map I h Filtering is performed to obtain the filtered heterogeneity mapping map M1;

[0025] Compactness Mapping Graph I using DoG2 kernel c Filtering is performed to obtain the filtered compactness map M2;

[0026] S23. Fuse the filtered heterogeneous mapping map M1 and the filtered compact mapping map M2 to obtain the fused target region feature image M. b

[0027] M b =M1×M2

[0028] Where × represents the multiplication of corresponding points in the pixel matrix.

[0029] Preferably, step three specifically includes:

[0030] S31. Constructing the fused target region feature image M based on the structure tensor. b Target corner features;

[0031] S32. Obtain the fused target region feature image M according to the matrix eigenvalue solving formula. b The two feature values ​​λ1 and λ2 at each pixel;

[0032] S33. Detect targets by calculating the region reflecting the characteristics of strong corner points using the angular intensity function:

[0033]

[0034] Where (x,y) represents the pixel coordinates, and λ1(x,y) and λ2(x,y) represent the two feature values ​​of the pixel (x,y) respectively;

[0035] S34. Introduce background suppression factor ω bp (x,y)

[0036] ω bp (x,y)=max(λ1(x,y),λ2(x,y))

[0037] S35, Based on the detection result ω of the angular intensity function cs (x,y) and background suppression factor ω bp (x,y), solve for the target response map M r

[0038] M r (x,y)=w cs (x,y)·w bp (x,y)

[0039] Where · represents the multiplication of corresponding coordinate values.

[0040] Preferably, step four specifically includes:

[0041] S41. Construct a sliding box W(p,q) with the point (p,q) as the center and a size of n×n;

[0042] The region centered at m×m (m<n) is defined as the mean-finding region, and the mean result obtained from the mean-finding region is expressed as v. mean (W(p,q));

[0043] The outermost annular region of the sliding frame W(p,q) is defined as the maximum value solution region, and the maximum value result obtained in the maximum value solution region is represented as v. max (W(p,q));

[0044] S42. Construct a target candidate region image M based on the mean and maximum values. c ,

[0045]

[0046] S43, the target candidate region image M c And target response map M r The data is fused to obtain the final target detection result M. d

[0047] M d =M r ×M c

[0048] Where × represents the multiplication of corresponding points in the image pixel matrix.

[0049] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method for detecting weak infrared targets by multiple sensing of target features, which has the following beneficial effects:

[0050] 1) This invention proposes a novel fast detection algorithm for weak targets in infrared single-frame images. It integrates and improves key operations in methods based on background consistency features, methods based on target saliency features, and methods based on low-rank and sparse decomposition of image structure, thus overcoming the shortcomings of single detection algorithms.

[0051] 2) This invention achieves a more robust detection algorithm by innovatively integrating target feature extraction strategies from multiple types of algorithms;

[0052] 3) Compared with other detection algorithms, the detection algorithm in this invention has good detection performance in terms of target enhancement and background suppression, and has high computational efficiency. Attached Figure Description

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0054] Figure 1 This is a schematic diagram of the method flow provided by the present invention;

[0055] Figure 2 This is a schematic diagram of the process framework provided in an embodiment of the present invention;

[0056] Figure 3 This is a schematic diagram illustrating the detection results provided in an embodiment of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] like Figure 1 As shown in the figure, this invention discloses a method for detecting weak infrared targets using multiple sensing of target features, comprising the following steps:

[0059] Step 1: Construct two new target region feature images using morphological erosion and dilation operations respectively;

[0060] Step 2: Use the DoG filter to filter the two new target region feature images, and fuse the filtering results to obtain the fused target region feature image;

[0061] Step 3: Use the structure tensor to extract corner features from the fused target region feature image to obtain the target response map;

[0062] Step 4: Use the target candidate region image to remove false targets in the target response map to obtain the target detection result.

[0063] Each step in the embodiments is described in detail below.

[0064] Step one is mainly used to construct new target features based on morphological combination operations.

[0065] In this step, morphological erosion and dilation operations are performed on the input infrared image I using the same structuring element SE to obtain the eroded image I. e and dilated image I d ,

[0066] I e =I⊙SE

[0067]

[0068] In this context, “⊙” and “⊕” represent the symbols for corrosion and expansion operations, respectively, and SE represents a structuring element with a radius of 1.

[0069] Morphological erosion involves sliding a structuring element (SE) across the image, extracting the minimum pixel value of the area it covers, and replacing the pixel at the center point (anchor point) of the structuring element. Similarly, morphological dilation involves sliding a structuring element (SE) across the image, extracting the maximum pixel value of the area it covers, and replacing the pixel at the anchor point. In infrared images, targets typically exhibit high reflectivity and high contrast, while the background is continuous and smooth. Generally, the center of the target is located at its highest pixel.

[0070] In erosion image I e In the process, the target area shrinks radially inward, thus expanding its surrounding background area. Although the overall intensity of the target will be slightly reduced, it still retains significant characteristics.

[0071] Using input infrared image I and erosion image I e The difference between corresponding pixels is used to obtain the compactness mapping I. c While smoothing the background and suppressing highlight noise, it can preserve the salient features of the target, thus constructing a feature with more significant local contrast in the target area, compactness mapping diagram I. c The calculation formula is

[0072] I c =II e .

[0073] In the dilated image I d In the process, the target area will expand radially outward, thus increasing the target's area and enhancing its overall brightness.

[0074] Using dilated image I d The heterogeneity mapping I is obtained by calculating the difference between corresponding pixels in the input infrared image I and the input infrared image I. h It can suppress continuous large-area background, and due to the input infrared image I and dilated image I d The pixel intensity of the target center point is consistent, therefore in the heterogeneity mapping map I h The pixel intensity at this point is 0, thus constructing a feature where the target center intensity is zero, and the surrounding area exhibits a ring of brightness. Heterogeneity mapping I h The calculation formula is

[0075] I h =I d -I.

[0076] The purpose of this step is to construct two new, easily perceptible target feature images based on the limited prior information (infrared weak targets of 2*2 to 9*9 pixels, which are point-like regions without shape or texture features, with pixel values ​​decreasing from the center outwards). The new target region feature images have the following characteristics:

[0077] 1) Compactness Mapping I c The local contrast in the target area is more significant;

[0078] 2) Heterogeneity Mapping I h The target center intensity is zero, and the surrounding background area is characterized by a ring of high brightness.

[0079] Use the two new target feature images mentioned above as the perceptual input for the next step.

[0080] Step two is mainly used to filter the two new feature images using the DoG filter.

[0081] DoG filters, as bandpass filters, are used to effectively extract the target region and suppress low-frequency background and high-frequency clutter noise. The formula for the DoG kernel can be expressed as:

[0082]

[0083] Where i and j represent the coordinate indices of the DoG kernel, and σ1 and σ2 represent the standard deviations of the two Gaussian functions, respectively. For a Gaussian function, the larger the σ, the smoother the peak; conversely, the smaller the σ, the more prominent the peak.

[0084] In this step, the purpose of using the DoG filter is to utilize two different DoG1 and DoG2 kernels to perceive the two new target features constructed in step one.

[0085] Since the size of small infrared targets is generally no larger than 9×9 pixels, the size of the two DoG cores is set to 5×5 pixels.

[0086] In heterogeneous mapping diagram I h In the model, the target center intensity is 0, and its surrounding area exhibits high brightness. Setting σ1 and σ2 of DoG1 to 2 and 0.5 respectively, the coefficients of the central surrounding region of the DoG1 kernel constructed based on the DoG kernel formula are larger than the coefficients of the center point. Figure 2 The 3D DoG1 is shown in Figure I. The heterogeneity mapping graph I is mapped using DoG1. h Filtering can extract target features that appear as a ring of bright background in the surrounding area.

[0087] In the compactness mapping diagram I c In this context, the target region is independent and highly significant. Setting σ1 and σ2 of DoG2 to 2 and 3 respectively, the coefficients of the DoG2 kernel constructed based on the DoG kernel formula are larger in the central region than in its surrounding regions. For example... Figure 2 The 3D DoG2 is shown in Figure I. The compactness mapping is performed using DoG2. c Filtering can extract target features with significant local contrast in the target region.

[0088] DoG1 and DoG2 respectively map the heterogeneity map I. h Compactness Mapping I c Filtering yields M1 and M2, enabling the perception of the two new target features constructed in step one. The two DoG filtering operations effectively extract the target region, enhance target contrast, and simultaneously suppress a significant amount of background clutter and noise, smoothing the bright background. Both filtering results retain some strong edge regions of the background; however, due to the opposite directions of erosion and dilation in the morphological operations, the same strong edge in the input infrared image I may appear in image I... c and I hIn M1 and M2, the edges in the original images do not overlap; similarly, the strong edges in the original images also do not overlap. Therefore, fusing M1 and M2 by multiplication yields M. b While preserving the saliency of the target, the influence of background clutter and noise is further suppressed, which can be specifically expressed as:

[0089] M b =M1×M2

[0090] Where × represents the Hadamard product of corresponding points in the pixel matrix of the image.

[0091] Step three primarily aims to extract corner features from the fused target region feature image to obtain the target response map.

[0092] Specifically, based on the fused target region feature image M obtained in step two... b Further perceptual extraction of the target's corner features is needed. Here, structure tensor information is used to construct the target's corner information. The formula for calculating the structure tensor J is:

[0093]

[0094] Among them G α This represents a Gaussian function with a standard deviation of α = 9, * indicates a convolution operation, and M... bx and M by These represent the gradients in the x and y directions of the feature image of the target region, respectively.

[0095] The formula for finding the eigenvalues ​​of a matrix is ​​expressed as:

[0096] Jz=λz

[0097] Where J is the structure tensor matrix of the eigenvalue to be solved, and z is the eigenvector of the structure tensor matrix J with respect to the eigenvalue λ.

[0098] The characteristic polynomial of the structure tensor matrix J can be obtained from the formula for solving matrix eigenvalues.

[0099] Recorded as:

[0100] |(λ)=|λE-J|, where E is the identity matrix.

[0101] After solving for the eigenvalues ​​of the second-order structure tensor matrix J, the fusion result M is obtained. b The relationship between the two feature values ​​λ1 and λ2 at each pixel of an image reflects the local geometric structure characteristics of the corresponding region. In flat regions, λ1≈λ2≈0; in edge regions, λ1>>λ2≈0; and in corner regions, λ1≥λ2>>0.

[0102] Small targets in infrared images often exhibit strong corner geometric characteristics, but target edges also objectively exist. Target detection can be achieved by calculating the angular intensity function to reflect the regions with strong corner characteristics.

[0103]

[0104] Where (x,y) represents the fusion result M b In the image, the pixel coordinates, λ1(x,y) and λ2(x,y) represent the two feature values ​​of the pixel (x,y), respectively.

[0105] In flat and edge regions, since λ2(x,y)≈0, then ω cs (x,y) approaches 0. In the corner region, since λ1(x,y)≥λ2(x,y)>>0, ω cs (x, y) will be a value greater than 0. To further suppress the influence of background noise, a background suppression factor ω is introduced. bp (x,y),

[0106] ω bp (x,y)=max(λ1(x,y),λ2(x,y))

[0107] In this step, based on the fusion result M b The structural tensor and eigenvalue information can be used to solve the target response map M. r This allows for a prominent target and a clean background.

[0108] M r (x,y)=w cs (x,y)·w bp (x,y)

[0109] Where · represents the product of corresponding values ​​at coordinate point (x, y).

[0110] Step four primarily aims to eliminate false targets using a target candidate mechanism.

[0111] This step will solve for the target candidate region based on the input infrared image I, in order to eliminate false targets outside the target candidate region generated in subsequent processing. The specific operation is as follows:

[0112] Construct a sliding bounding box W(p,q) centered at point (p,q) with a size of n×n. Define the region centered at m×m (m<n) as the mean calculation region, and represent the mean result as v. mean (W(p,q)), the outermost annular region of W(p,q) is defined as the maximum value solution region, and the maximum value result is expressed as v. max (W(p,q)). Combining the heterogeneity and compactness characteristics of infrared weak targets, it can be seen that when vmean (W(p,q))>v max When W(p,q) is true, the central region of the window has a bright, salient feature, while its surrounding neighborhood has a lower pixel value. Therefore, this region is defined as the target candidate region. Conversely, if the salientity of the central region of the window is lower than that of its surrounding neighborhood, it is defined as the background region. Considering the size range of small targets (2×2 to 9×9) and their high salientity at the center, this invention sets m to 3 and n to 7. This ensures effective identification of weak targets while minimizing interference from the bright background around the target and eliminating randomly generated clutter and noise. The target candidate region image obtained by solving is defined as M. c This is a binary mask image.

[0113]

[0114] For a real target, its central region has a prominent, bright feature, while the surrounding background has a lower pixel value. When W(p,q) slides to this region, the mean calculation area within the sliding window contains the main part of the target, while the maximum value calculation area is the background pixels. Therefore, this region can be marked as a candidate region for the target (i.e., the region with a pixel value of 1 in the binary image) using the above formula. For pseudo-targets such as clutter noise, their distribution is random, their size is smaller, and their salience is relatively lower than that of real targets. When W(p,q) slides to this region, as shown in the above formula, the mean calculation area within the sliding window contains a large amount of background information, while the maximum value calculation area usually contains other clutter noise or a bright background. Therefore, v mean (W(p,q)) is almost smaller than v max (W(p,q)), thus marking the region as a pseudo-target region (i.e., a region with a pixel value of 0 in a binary image).

[0115] The calculated binarized target candidate region image M c This can eliminate the vast majority of false target regions, and then combine it with the target response map M obtained through multiple target feature perception in the previous steps. r This enables effective detection of weak infrared targets. The false target removal operation involves processing the target response map M... r With the target candidate region image M c The data is fused to obtain the final target detection result M. d :

[0116] M d =M r ×M c

[0117] Where × represents the multiplication of corresponding points in the image pixel matrix.

[0118] To verify the effectiveness of this invention, several common infrared images were used as input, and the invention was used to detect weak infrared targets. Figure 3 As shown in the figure, the first row is the input image, the second row is the detection result of the present invention, and the box marks the real target area.

[0119] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0120] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for detecting weak infrared targets using multiple sensing of target features, characterized in that, The method includes the following steps: Step 1: Construct two new target region feature images using morphological erosion and dilation operations respectively; specifically including: S11, process the input infrared images respectively Morphological erosion and dilation operations are performed to obtain the eroded image. and dilated image ; S12, Utilizing the input infrared image With corrosion images The difference value is used to obtain a compactness mapping. : ； Using dilated images With input infrared image The difference value is used to obtain a heterogeneity mapping graph. : I； The compact mapping diagram Heterogeneity mapping diagram These are two new target region feature images; Step 2: Filter the two new target region feature images using the DoG filter, and fuse the filtering results to obtain the fused target region feature image; specifically including: S21. Based on the DoG filter, obtain two DoG cores with different standard deviations, DoG1 core and DoG2 core, wherein the DoG1 core contains... and The values ​​are set to 2 and 0.5 respectively, in the DoG2 core. and Set them to 2 and 3 respectively; S22. Use the DoG1 kernel to map heterogeneity. Filtering is performed to obtain the filtered heterogeneity map. ; Compactness mapping using DoG2 cores Filtering is performed to obtain the filtered compactness map. ; S23. Filter the heterogeneity mapping map. Compactness mapping after filtering The fusion process is performed to obtain the feature image of the fused target region. : ； in, This represents the product of corresponding points in the pixel matrix; Step 3: Extract corner features from the fused target region feature image using the structure tensor to obtain the target response map; specifically including: S31. Constructing the fused target region feature image based on the structure tensor. Target corner features; S32. Obtain the fused target region feature image based on the matrix eigenvalue solving formula. Two feature values ​​at each pixel and ; S33. Detect targets by calculating the region reflecting the characteristics of strong corner points using the angular intensity function: ； in Represents pixel coordinates, and Representing pixels Two eigenvalues; S34. Introduce background suppression factor : ； S35. Detection results based on angular intensity function and background inhibition factor Solve the target response map : ； in, This represents the product of corresponding coordinate values; Step 4: Use the target candidate region image to remove false targets in the target response map to obtain the target detection result; specifically including: S41. Construct a point-based system. Centered on, size is sliding box ; center ( The region is defined as the mean-finding region, and the mean result obtained from solving the mean-finding region is expressed as follows: ; slide box The outermost annular region is defined as the maximum value solution region, and the maximum value result obtained in the maximum value solution region is expressed as follows: ; S42. Construct a target candidate region image based on the mean and maximum values. , ； S43, the target candidate region image and target response diagram The data is fused to obtain the final target detection result. : ； in, This represents the product of corresponding points in the image pixel matrix.

2. The infrared weak target detection method based on multiple sensing of target features according to claim 1, characterized in that, In step S11, the input infrared images are processed respectively. Morphological erosion and dilation operations are performed to obtain the eroded image. and dilated image Specifically, it includes the following formulas: ； ； in" "and" "These represent the symbols for corrosion and expansion operations, respectively." This represents a structuring element with a radius of 1.

Images