Mid-infrared Dim and Small Target Detection Method and System with Multi-Feature Fusion in Complex Background

Abstract

A method and system for detecting small and weak infrared targets in complex backgrounds with multi-feature fusion belong to the technical field of infrared target detection and recognition, and solve the problem that the existing technology has poor algorithm adaptability when detecting small and weak infrared targets in complex backgrounds, resulting in a large number of false alarms in the detection results and affecting the accuracy of the detection results. The present invention first extracts radiation features, multi-order directional derivative features and spectral features representing the radiation characteristics, structural characteristics and intensity characteristics of small and weak targets respectively, fuses multiple features to construct a feature saliency map, enhances the target while suppressing background noise; then uses the CFAR adaptive detection method to calculate the segmentation threshold of the image, obtains a binary segmentation result, and performs morphological processing to screen out false targets caused by isolated points and noise, and obtains the final detection result of small and weak targets. The algorithm of the present invention has low complexity, strong adaptability to complex backgrounds, high detection accuracy and is convenient for engineering implementation.

Description

Technical Field

[0001] The present invention belongs to the technical field of infrared target detection and recognition, and relates to a method and system for detecting infrared small and weak targets in complex backgrounds with multi-feature fusion. Background Art

[0002] For the traditional single-feature threshold segmentation detection algorithm for small and weak targets in infrared images, the accuracy of its detection results highly depends on the target pixel intensity and the number of pixels. Since the pixel ratio of small targets in infrared images is very low, small targets are often submerged by the background and noise. Therefore, for infrared images, directly using the preset threshold segmentation method for target detection has a relatively single processing feature, resulting in a low target detection rate and a high false alarm rate.

[0003] At present, some studies on infrared target detection methods based on multi-feature fusion have been carried out at home and abroad, and available algorithm models have been formed, but several important problems have not been solved yet, mainly manifested in: First, most of the current infrared target detection algorithms based on multi-feature fusion mainly focus on targets above medium size, and have poor feature extraction ability for targets with low "target-background" contrast and small number of pixels, making it difficult to obtain a high target detection rate. Second, the existing algorithms have poor adaptability in detecting infrared small and weak targets in complex backgrounds, resulting in a large number of false alarms in the detection results and affecting the accuracy of the detection results.

[0004] The literature "Research on Infrared Small Target Detection Algorithm under Complex Background" (Tian Wei, Northeastern University) with a publication date of June 2012 analyzed the three elements of targets, background, and noise in infrared images around the problem of infrared small and weak target detection. By analyzing infrared small and weak target images, a quantitative description of the regional complexity of infrared images was extracted, and the variance-weighted information entropy, gradient direction feature, and local contrast feature of the images were analyzed and discussed. Since it is difficult to ensure the detection results with single-frame detection, a sequence image detection algorithm based on three-dimensional wavelet transform was proposed. However, the technical solution of the above literature still has problems such as poor algorithm adaptability, low accuracy, resulting in a large number of false alarms in the detection results and affecting the accuracy of the detection results when detecting infrared small and weak targets in complex backgrounds. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for detecting infrared small and weak targets in complex backgrounds with multi-feature fusion, so as to solve the problems in the prior art that the algorithm has poor adaptability in detecting infrared small and weak targets in complex backgrounds, resulting in a large number of false alarms in the detection results and affecting the accuracy of the detection results.

[0006] Therefore, a technology for detecting infrared small and weak targets in complex backgrounds with multi-feature fusion is proposed to increase the interpretability of small and weak targets in infrared images and improve the accuracy of detection results.

[0007] The present invention solves the above technical problems through the following technical solutions:

[0008] A method for detecting small and weak infrared targets in a complex background with multi-feature fusion, comprising the following steps:

[0009] S1. Input the original infrared small target image, preprocess the input image using a high-pass filter, and extract the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map of the image respectively;

[0010] S2. Adopt a normalized feature fusion method based on the visual attention mechanism to fuse the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map to generate a saliency feature fusion map;

[0011] S3. For the saliency feature fusion map, first use the CFAR algorithm to traverse each pixel gray value and judge whether it exceeds the adaptive segmentation threshold to achieve the detection of small and weak targets, then use the pixel clustering method to cluster the detection results, and finally screen out isolated points and false targets caused by noise to obtain the target segmentation result.

[0012] The technical solution proposed by the present invention first extracts the radiation feature, multi-order directional derivative feature, and spectral feature representing the radiation characteristics, structural characteristics, and intensity characteristics of small and weak targets respectively, and fuses multiple features to construct a feature saliency map, enhancing the target while suppressing background noise; then uses the CFAR adaptive detection method to calculate the segmentation threshold of the image, obtains a binary segmentation result, and performs morphological processing to screen out isolated points and false targets caused by noise to obtain the final small and weak target detection result; the algorithm of the present invention has low complexity, strong adaptability to complex backgrounds, high detection accuracy, and is convenient for engineering implementation.

[0013] As a further improvement of the technical solution of the present invention, the extraction method of the radiation feature saliency map in step S1 is as follows:

[0014] 1) Modify the original infrared small target image through environmental impact correction to obtain a two-dimensional temperature field distribution map, and the correction formula is as follows:

[0015] B(T obs ) = B(T target )·ε·τ atm + B(T atm )·(1 - τ atm )

[0016] 2) According to the obtained two-dimensional temperature field distribution map, calculate the two-dimensional distribution map of the zero-line-of-sight blackbody equivalent bright temperature as follows:

[0017]

[0018] 3) The two-dimensional distribution map of the equivalent bright temperature of the zero viewing distance black body is inversely Planck-transformed to obtain the inversion result T of the zero viewing distance temperature field target , that is, the radiation feature significant map:

[0019] S radiation = T target = B -1 (L)

[0020] where B(T obs ) is the black body radiation with the temperature equal to T obs , B(T target ) is the black body radiation temperature corresponding to the target true temperature T target , ε is the equivalent emissivity of the black body within the thermal imager band, B(T atm )·(1 - τ atm ) is the path radiation superimposed on the observed value, and the parameter τ atm is related to the environmental parameters.

[0021] As a further improvement of the technical solution of the present invention, the extraction method of the multi-order directional derivative feature significant map in step S1 is as follows:

[0022] a) Extract the multi-order directional derivative features of each pixel point (x0, y0) on the original infrared small target image along the direction vector as follows:

[0023]

[0024] b) Calculate according to the least squares surface fitting and the orthogonality of polynomials:

[0025]

[0026] to obtain:

[0027]

[0028]

[0029]

[0030] Thus, three weight coefficient matrices are obtained as follows:

[0031]

[0032]

[0033]

[0034] where α is the angle between the direction vector and the x-axis, and β is the direction vector The included angle between the y-axis, K i (i = 4, 5, 6) represents the weight coefficient; where the parameters r and c respectively represent the buffer radii in the x-axis and y-axis directions, I(x + r, y + c) represents the pixel value at the point (x + r, y + c), P i (r, c) represents the direction vector of the point (r, c);

[0035] c) Modify the obtained multi-order directional derivative feature map, traverse the pixel values in the feature map, set them to zero if they are greater than zero, then normalize the feature map, and use a 3×3 filtering window to globally process the image. Cross-fuse the images on each direction channel, perform dot multiplication on the mutually orthogonal feature map vectors, suppress background clutter noise, and enhance weak targets to obtain the multi-order directional derivative feature significant map as follows:

[0036]

[0037] Among them, g represents the number of groups of orthogonal bases; S g represents the direction feature map, ⊥S g represents the direction feature map orthogonal to S g , and N(·) represents the normalization function.

[0038] As a further improvement of the technical solution of the present invention, the extraction method of the spectral feature significant map in step S1 is:

[0039] Ⅰ) In the frequency domain, perform Fourier transform on the preprocessed original infrared small target image to obtain:

[0040] I F = F(I orig )

[0041] Ⅱ) Separate the amplitude spectrum and phase spectrum of the original data from the spectral image to obtain:

[0042] A f = Abs(I F )

[0043] P f = Angle(I F )

[0044] Ⅲ) Use the mean filter h n (f) to fit the background amplitude spectrum of the image, and the template size is adaptively adjusted according to the target size in the image domain to obtain:

[0045] L(f) = log(A f )

[0046] L f_smooth = h n (f)*Lf

[0047] Among them, L f is the image after taking the logarithm of the amplitude spectrum information, and L f_smooth is the result after mean filtering on this basis, and h n (f) is the template matrix, and the calculation formula of h n (f) is as follows:

[0048]

[0049] Ⅳ) Remove the estimated background part to obtain the significant log amplitude spectrum:

[0050] R f = L f - L f_smooth

[0051] Ⅴ) Add the significant log amplitude spectrum to the phase spectrum in step Ⅱ) and perform IFFT transformation and high-pass filtering enhancement to obtain the spectral feature significant map:

[0052] S x = g x * F -1 [exp(R f + P f )] 2 .

[0053] As a further improvement of the technical solution of the present invention, the calculation formula of the significant feature fusion map described in step S2 is:

[0054]

[0055] Among them, S radiation , S MODD and S x respectively represent the radiation feature significant map, the multi-order directional derivative feature significant map, and the spectral feature significant map.

[0056] As a further improvement of the technical solution of the present invention, the calculation formula of the adaptive segmentation threshold described in step S3 is:

[0057]

[0058] Among them, T is the adaptive segmentation threshold, P fa is the set false alarm rate, and p(x) is the background probability density distribution function.

[0059] The mid-infrared small and weak target detection system with multi-feature fusion in complex backgrounds includes: a feature significant map extraction module, a significant feature fusion module, and a detection result output module,

[0060] The described feature saliency map extraction module: Input the original infrared small target image, preprocess the input image using a high-pass filter, and extract the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map of the image respectively;

[0061] The described saliency feature fusion module uses a normalized feature fusion method based on the visual attention mechanism to fuse the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map to generate a saliency feature fusion map;

[0062] The described detection result output module: For the saliency feature fusion map, first use the CFAR algorithm to traverse each pixel gray value and determine whether it exceeds the adaptive segmentation threshold to detect small and weak targets, then use the pixel clustering method to cluster the detection results, and finally screen out isolated points and false targets caused by noise to obtain the target segmentation result.

[0063] As a further improvement of the technical solution of the present invention, the extraction method of the radiation feature saliency map is as follows:

[0064] 1) Modify the original infrared small target image through environmental impact correction to obtain a two-dimensional temperature field distribution map. The correction formula is as follows:

[0065] B(T obs )=B(T target )·ε·τ atm +B(T atm )·(1-τ atm )

[0066] 2) According to the obtained two-dimensional temperature field distribution map, calculate the two-dimensional distribution map of the zero-line-of-sight blackbody equivalent bright temperature as follows:

[0067]

[0068] 3) Perform inverse Planck transformation on the two-dimensional distribution map of the zero-line-of-sight blackbody equivalent bright temperature to obtain the zero-line-of-sight temperature field inversion result T target , that is, the radiation feature saliency map:

[0069] S radiation =T target =B -1 (L)

[0070] Where, B(T obs ) is the blackbody radiation with a temperature equal to T obs , B(T target ) is the blackbody radiation temperature corresponding to the target true temperature T target , ε is the blackbody equivalent emissivity within the thermal imager band, B(T atm )·(1-τ atm) is the path radiance superimposed on the observed value, and the parameter τ atm is related to the environmental parameters;

[0071] The method for extracting the multi-order directional derivative feature map is as follows:

[0072] a) Extract the multi-order directional derivative features of each pixel point (x0, y0) on the original infrared small target image along the direction vector as follows:

[0073]

[0074] b) Calculate according to the least squares surface fitting and the orthogonality of polynomials:

[0075]

[0076] to obtain:

[0077]

[0078]

[0079]

[0080] Thus, three weight coefficient matrices are obtained as follows:

[0081]

[0082]

[0083]

[0084] where α is the angle between the direction vector and the x-axis, β is the angle between the direction vector and the y-axis, and K i (i = 4, 5, 6) represents the weight coefficient; where the parameters r and c respectively represent the buffer radii in the x-axis and y-axis directions, I(x + r, y + c) represents the pixel value at the point (x + r, y + c), and P i (r, c) represents the direction vector of the point (r, c);

[0085] c) Correct the obtained multi-order directional derivative feature map, traverse the pixel values in the feature map, set them to zero if they are greater than zero, then normalize the feature map, and use a 3×3 filtering window to perform global processing on the image. Cross-fuse the images on each direction channel, perform dot multiplication on the mutually orthogonal feature map vectors, suppress background clutter noise, and enhance weak small targets to obtain the multi-order directional derivative feature significant map as follows:

[0086]

[0087] Among them, g represents the number of groups of orthogonal bases; S g represents the directional feature map, ⊥S g represents the directional feature map orthogonal to S g , and N(·) represents the normalization function;

[0088] The method for extracting the spectral feature saliency map is as follows:

[0089] Ⅰ) In the frequency domain, perform Fourier transform on the preprocessed original infrared small target image to obtain:

[0090] I F = F(I orig )

[0091] Ⅱ) Separate the amplitude spectrum and phase spectrum of the original data from the spectral image to obtain:

[0092] A f = Abs(I F )

[0093] P f = Angle(I F )

[0094] Ⅲ) Use the mean filter h n (f) to fit the background amplitude spectrum of the image, and the template size is adaptively adjusted according to the target size in the image domain to obtain:

[0095] L(f) = log(A f )

[0096] L f_smooth = h n (f)*L f

[0097] Among them, L f is the image after taking the log of the amplitude spectrum information, L f_smooth is the result after mean filtering on this basis, h n (f) is the template matrix, and the calculation formula of h n (f) is as follows:

[0098]

[0099] Ⅳ) Remove the estimated background part to obtain the significant log amplitude spectrum:

[0100] R f = L f - L f_smooth

[0101] Ⅴ) Add the significant log magnitude spectrum to the phase spectrum in step Ⅱ), perform IFFT transformation and high-pass filtering enhancement, and then the spectrum feature significant map can be obtained:

[0102] S x = g x * F -1 [exp(R f + P f )] 2 .

[0103] As a further improvement of the technical solution of the present invention, the calculation formula of the significant feature fusion map is:

[0104]

[0105] Wherein, S radiation , S MODD and S x respectively represent the radiation feature significant map, the multi-order directional derivative feature significant map, and the spectrum feature significant map.

[0106] As a further improvement of the technical solution of the present invention, the calculation formula of the adaptive segmentation threshold is:

[0107]

[0108] Wherein, T is the adaptive segmentation threshold, P fa is the set false alarm rate, and p(x) is the background probability density distribution function.

[0109] The advantages of the present invention are as follows:

[0110] The technical solution proposed by the present invention first extracts the radiation feature, multi-order directional derivative feature, and spectrum feature that respectively represent the radiation characteristics, structural characteristics, and intensity characteristics of small and weak targets, fuses multiple features, constructs a feature significant map, enhances the target while suppressing background noise; then uses the CFAR adaptive detection method to calculate the segmentation threshold of the image, obtains a binary segmentation result, and performs morphological processing to screen out false targets caused by isolated points and noise, and finally obtains the detection result of small and weak targets; the algorithm complexity of the present invention is low, it has strong adaptability to complex backgrounds, and the detection accuracy is relatively high, which is convenient for engineering implementation. Description of the Drawings

[0111] Figure 1 is the flowchart of the method for detecting small and weak infrared targets in a complex background with multi-feature fusion according to Embodiment 1 of the present invention;

[0112] Figure 2 is the flowchart for extracting the spectrum feature significant map according to Embodiment 1 of the present invention;

[0113] Figure 3It is a schematic diagram of weak target detection in the first embodiment of the present invention. Detailed implementation manners

[0114] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0115] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings of the specification and specific embodiments:

[0116] Embodiment 1

[0117] As Figure 1 shown, a method for detecting infrared weak targets in a complex background with multi-feature fusion includes the following steps:

[0118] Step 1: Perform environmental impact correction on the original infrared digital image through the following calculation formula to obtain a two-dimensional distribution map of the zero-line-of-sight blackbody equivalent bright temperature:

[0119]

[0120] B(T obs ) is the blackbody radiation at a temperature equal to T obs , that is, the observed bright temperature of the thermal imager, which can be obtained according to the instrument observation results; B(T target ) is the blackbody radiation temperature corresponding to the target true temperature T target , obtained from the instrument parameter table; ε is the blackbody equivalent emissivity within the thermal imager band; B(T atm )·(1 - τ atm ) is the path radiation superimposed on the observed value, and τ a m is related to environmental parameters such as the observation distance, atmospheric temperature and humidity.

[0121] After inverse Planck transformation, the inversion result T target of the zero-line-of-sight temperature field can be obtained, that is, the radiation feature is significant S radiation :

[0122] S radiation = T target = B -1 (L)

[0123] Step 2: On the basis of the quantized original infrared digital image, extract the multi-order directional derivative features of each pixel point (x0, y0) on the image along the direction vector :

[0124]

[0125] α is the angle between the vector and the x-axis, and β is the angle between and the y-axis, and K i (i = 4, 5, 6) The coefficients are related to the pixel coordinates of the input image, denoted as K i (x, y) (i = 4, 5, 6), calculated according to the least squares surface fitting and the orthogonality of polynomials:

[0126]

[0127] The calculation results are:

[0128]

[0129]

[0130]

[0131] Calculate three weight templates:

[0132]

[0133]

[0134]

[0135] And it is necessary to correct the obtained direction feature map. Traverse the pixel values in the direction feature map. If it is greater than zero, set it to zero; perform normalization processing on the image, and the value range is [0, 1]; use a 3×3 filtering window to perform global processing on the image.

[0136] For the images on each direction channel, perform cross-fusion to obtain the direction feature map:

[0137]

[0138] Among them, g represents the number of groups of orthogonal bases; S g represents the direction feature map, ⊥S g represents the direction feature map orthogonal to S g N(·) represents normalization, and S MODD represents the fused feature map. Perform a dot product on the mutually orthogonal feature map vectors.

[0139] Step 3. Based on the quantized original infrared digital image, extract each pixel point (x0, y0) on the image along the direction vector The multi-order directional derivative feature is obtained, that is, the saliency map of the image is obtained. In the frequency domain, the original image is Fourier-transformed:

[0140] I F = F(I orig )

[0141] The amplitude spectrum information and phase spectrum information of the original data are separated from the spectral image:

[0142] A f = Abs(I F )

[0143] P f = Angle(I F )

[0144] The background amplitude spectrum of the image is fitted with a mean filter h n (f), and the template size can be adaptively adjusted according to the target size in the image domain.

[0145] L(f) = log(A f )

[0146] L f_smooth = h n (f) * L f

[0147] where L f is the image after taking the log of the amplitude spectrum information, and L f_smooth is the result after mean filtering on this basis. h n (f) is the template:

[0148]

[0149] The estimated background part is removed, that is, L f_smooth , and the remaining part is the log amplitude spectrum of the saliency. Adding the previously extracted phase spectrum for IFFT transformation and high-pass filtering enhancement:

[0150] R f = L f - L f_smooth

[0151] S x = g x * F -1 [exp(R f + P f )] 2

[0152] Step 4: In the above saliency map, according to the visual attention mechanism, enhance its saliency. For the radiation feature saliency map S radiation , the multi-order directional derivative feature saliency map SMODD 、Spectral feature significant graph S x Fuse the three feature graphs according to the fusion algorithm to obtain the final significant feature fusion graph S FSM :

[0153]

[0154] Step 5. For the significant feature fusion graph S FSM , use the CFAR method to determine the background distribution model by traversing the gray level of each pixel in the global image, estimate the parameters of the background probability density distribution function p(x), and set the false alarm rate P fa , solve the adaptive segmentation threshold T according to the formula:

[0155]

[0156] Judge pixel by pixel whether it exceeds the adaptive segmentation threshold to detect small and weak targets. Then use the pixel clustering method to cluster the detection results, and finally screen out isolated points and false targets caused by noise to obtain the target segmentation result.

[0157] Aiming at the problem of detecting small and weak targets in complex backgrounds in infrared images, the present invention proposes an infrared small and weak target detection technology in complex backgrounds based on multi-feature fusion. After obtaining the original infrared image data, first, an improved high-pass filter is used for enhanced preprocessing. Then, the zero-line-of-sight temperature field radiation feature, multi-order directional derivative (MODD) feature, and spectral feature of the infrared small target are extracted respectively. A normalization method based on the visual attention mechanism is used to fuse the three types of features to generate a significant feature fusion image, strengthen the target saliency, suppress the background, and improve the detection ability of infrared small and weak signal targets in complex natural backgrounds. And based on the CFAR algorithm, the probability distribution of the image background is fitted and an adaptive segmentation threshold segmentation is performed to realize the detection of small and weak targets after significant enhancement on the feature fusion image. Feature fusion mainly uses the top-down visual attention mechanism. After successively obtaining the significant maps of the zero-line-of-sight temperature field inversion radiation feature, multi-order directional derivative feature, and spectral feature of the infrared small target, the generation of the fused feature significant map is completed by calculating the weighted geometric mean. The multi-feature fusion utilizes the visual saliency advantages of each feature map, extracts the best performance of each significant map, suppresses background noise and false alarm information, and projects the attention on the significant region of interest where the small and weak target is located in the fused image, avoiding the later detection task from focusing on the entire image, thereby reducing the false alarm rate of target detection and improving the detection performance. The entire multi-feature fusion infrared small and weak target detection algorithm module covers three parts: the first part is to preprocess the original image and extract the significant maps of the zero-line-of-sight temperature field radiation feature, multi-order directional derivative feature, and spectral feature of the target; the second part is to perform top-down multi-feature fusion on the above three feature significant maps based on the visual attention mechanism to generate a significant feature fusion image to suppress complex background noise, improve the signal-to-clutter ratio of the image, and improve the characterization ability of target characteristics; the third part is to use the CFAR detection method to perform adaptive segmentation threshold segmentation on the generated fusion image to detect the target pixels. Compared with the traditional single-feature threshold segmentation detection method for small and weak targets in infrared images, an infrared small and weak target detection technology in complex backgrounds based on multi-feature fusion does not require presetting a segmentation threshold, but based on fusing multiple types of features of small and weak targets, extracts the significant region of the target, and realizes the detection of small and weak targets through the CFAR adaptive segmentation threshold detection method.

[0158] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting small and weak mid-infrared targets in complex backgrounds with multi-feature fusion, characterized in that, It includes the following steps: S1. Input the original infrared small target image, preprocess the input image using a high-pass filter, and extract the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map of the image respectively; The extraction method of the multi-order directional derivative feature saliency map is as follows: a) Extract the pixel points on the original infrared small target image along the direction vector of the multi-order directional derivative feature; b) Calculate according to least squares surface fitting and the orthogonality of polynomials; c) Correct the obtained multi-order directional derivative feature map, traverse the pixel values in the feature map, set them to zero if they are greater than zero, then normalize the feature map, and globally process the image using a 3×3 filtering window, cross-fuse the images on each direction channel, perform dot product on the mutually orthogonal feature map vectors, suppress background clutter noise, and enhance weak small targets to obtain the multi-order directional derivative feature saliency map; The multi-order directional derivative feature saliency map is as follows: where g represents the number of groups of orthogonal bases; represents the directional feature map, represents the one orthogonal to the directional feature map, represents the normalization function; S2. Adopt a normalized feature fusion method based on the visual attention mechanism to fuse the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map to generate a saliency feature fusion map; S3. For the saliency feature fusion map, first use the CFAR algorithm to detect weak small targets by traversing each pixel gray value and judging whether it exceeds the adaptive segmentation threshold, then use the pixel clustering method to cluster the detection results, and finally screen out isolated points and false targets caused by noise to obtain the target segmentation result.

2. The method for detecting small and weak mid-infrared targets in complex backgrounds with multi-feature fusion according to claim 1, characterized in that, The extraction method of the radiation feature saliency map described in step S1 is as follows: 1) Modify the original infrared small target image through environmental impact correction to obtain a two-dimensional temperature field distribution map. The correction formula is as follows: 2) Calculate the two-dimensional distribution map of the zero-range blackbody equivalent bright temperature according to the obtained two-dimensional temperature field distribution map as follows: 3) The two-dimensional distribution map of the equivalent bright temperature of the zero viewing distance black body is inversely Planck-transformed to obtain the inversion result of the zero viewing distance temperature field , that is, the radiation characteristic significant map: Among them, is the blackbody radiation at a temperature equal to , is the blackbody radiation temperature corresponding to the target true temperature , is the blackbody equivalent emissivity within the thermal imager band, is the path radiation superimposed on the observed value, and the parameter is related to the environmental parameters.

3. The method for detecting small and weak mid-infrared targets in complex backgrounds with multi-feature fusion according to claim 2, characterized in that, a) Extract each pixel point on the original infrared small target image Along the direction vector The multi-order directional derivative features are as follows: b) The specific calculation according to least squares surface fitting and the orthogonality of polynomials is as follows: Obtain: Thus, the following three weight coefficient matrices are obtained: wherein, is the angle between the direction vector and the x-axis, is the angle between the direction vector and the y-axis, represents a weight coefficient; wherein the parameters r and c respectively represent the buffer radii in the x-axis and y-axis directions, represents the pixel value at the point (x + r, y + c), represents the direction vector of the point (r, c).

4. The method for detecting small and weak mid-infrared targets in complex backgrounds with multi-feature fusion according to claim 3, characterized in that, The extraction method of the spectral feature saliency map described in step S1 is as follows: Ⅰ) In the frequency domain, perform Fourier transform on the preprocessed original infrared small target image to obtain: Ⅱ) Separate the amplitude spectrum and phase spectrum of the original data from the spectral image to obtain: Ⅲ) Using a mean filter to fit the background amplitude spectrum of the image, and the template size is adaptively adjusted according to the target size in the image domain as follows: Among them, is the image after taking the logarithm of the amplitude spectrum information, is the result after mean filtering on this basis, is the template matrix, The calculation formula of is as follows: Ⅳ) Remove the estimated part in the background to obtain the significant log amplitude spectrum: Ⅴ) Add the significant log amplitude spectrum to the phase spectrum in step Ⅱ) and perform IFFT transform and high-pass filter enhancement to obtain the spectral feature saliency map: 。 5. The method for detecting a dim and small infrared target in a complex background with multi-feature fusion according to claim 4, wherein, The calculation formula of the saliency feature fusion map described in step S2 is: Among them, , and respectively represent the radiation feature significant map, the multi-order directional derivative feature significant map, and the spectral feature significant map.

6. The method for detecting a dim and small infrared target in a complex background with multi-feature fusion according to claim 5, wherein, The calculation formula of the adaptive segmentation threshold described in step S3 is: wherein, is the adaptive segmentation threshold, is the set false alarm rate, is the background probability density distribution function.

7. A system for detecting a dim and small infrared target in a complex background with multi-feature fusion, wherein, It includes: A feature saliency map extraction module, a saliency feature fusion module, and a detection result output module. The feature saliency map extraction module: Input the original infrared small target image, preprocess the input image using a high-pass filter, and extract the radiation feature saliency map, multi-order directional derivative feature saliency map, and spectral feature saliency map of the image respectively; The extraction method of the multi-order directional derivative feature saliency map is as follows: a) Extract the pixel points on the original infrared small target image along the direction vector of the multi-order directional derivative feature; b) Calculate according to least squares surface fitting and the orthogonality of polynomials; c) Modify the obtained multi-order directional derivative feature map. Traverse the pixel values in the feature map. If the value is greater than zero, set it to zero. Then, perform normalization processing on the feature map, and use a 3×3 filtering window to globally process the image. Cross-fuse the images on each direction channel, perform dot product on the mutually orthogonal feature map vectors, suppress background clutter noise, and enhance weak small targets to obtain a multi-order directional derivative feature saliency map; The multi-order directional derivative feature saliency map is as follows: where \(g\) represents the number of groups of orthogonal bases; represents the directional feature map, represents the one orthogonal to the directional feature map, represents the normalization function; The saliency feature fusion module adopts a normalization feature fusion method based on the visual attention mechanism to fuse the radiation feature saliency map, the multi-order directional derivative feature saliency map, and the spectral feature saliency map to generate a saliency feature fusion map; The detection result output module: For the saliency feature fusion map, first use the CFAR algorithm to traverse each pixel gray value and judge whether it exceeds the adaptive segmentation threshold to detect weak small targets, then use the pixel clustering method to cluster the detection results, and finally filter out isolated points and false targets caused by noise to obtain the target segmentation result.

8. The system for detecting a dim and small infrared target in a complex background with multi-feature fusion according to claim 7, wherein, The extraction method of the radiation feature saliency map is: 1) Modify the original infrared small target image through environmental impact correction to obtain a two-dimensional temperature field distribution map. The correction formula is as follows: 2) According to the obtained two-dimensional temperature field distribution map, calculate the two-dimensional distribution map of the zero-line-of-sight blackbody equivalent bright temperature as follows: 3) The two-dimensional distribution map of the equivalent bright temperature of the zero viewing distance black body is inversely Planck-transformed to obtain the inversion result of the zero viewing distance temperature field , that is, the radiation characteristic significant map: wherein, is the blackbody radiation at a temperature equal to , is the blackbody radiation temperature corresponding to the target true temperature , is the blackbody equivalent emissivity within the thermal imager band, is the path radiation superimposed on the observed value, and the parameter is related to the environmental parameters; a) Extract each pixel point on the original infrared small target image along the direction vector The multi-order directional derivative features are as follows: b) Calculate specifically according to the least squares surface fitting and the orthogonality of polynomials as follows: Obtain: Thus, obtain the following three weight coefficient matrices: wherein, is the angle between the direction vector and the x-axis, is the angle between the direction vector and the y-axis, represents a weight coefficient; wherein the parameters r and c respectively represent the buffer radii in the x-axis and y-axis directions, represents the pixel value at the point (x + r, y + c), represents the direction vector of the point (r, c); The extraction method of the spectral feature saliency map is: Ⅰ) In the frequency domain, perform Fourier transform on the preprocessed original infrared small target image to obtain: Ⅱ) Separate the amplitude spectrum and phase spectrum of the original data from the spectral image to obtain: Ⅲ) Use a mean filter to fit the background amplitude spectrum of the image, and the template size is adaptively adjusted according to the target size in the image domain as follows: Among them, is the image after taking the logarithm of the amplitude spectrum information, is the result after mean filtering on this basis, is the template matrix, The calculation formula of is as follows: Ⅳ) Remove the estimated part in the background to obtain the significant log amplitude spectrum: Ⅴ) Add the significant log amplitude spectrum to the phase spectrum in step Ⅱ) and perform IFFT transform and high-pass filtering enhancement to obtain the spectral feature saliency map: 。 9. The multi-feature fusion infrared small target detection system in complex background according to claim 8, characterized in that, The calculation formula of the saliency feature fusion map is: Among them, , and respectively represent the radiation feature significant map, the multi-order directional derivative feature significant map, and the spectral feature significant map.

10. The multi-feature fusion infrared small target detection system in complex background according to claim 9, characterized in that, The calculation formula of the adaptive segmentation threshold is: wherein, is the adaptive segmentation threshold, is the set false alarm rate, is the background probability density distribution function.

Images