A method for infrared small target detection based on multi-directional local difference measurement and entropy weighting

Through the multi-directional local difference measurement and entropy weighting method, the problem of high error detection in infrared small object detection is solved, and the performance of infrared small object detection in complex backgrounds is improved, significantly reducing the error detection rate and enhancing the target detection effect. .

CN114494903BActive Publication Date: 2024-11-01HUAZHONG NORMAL UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202210069992.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-01-21
Publication Date
2024-11-01
Estimated Expiration
2042-01-21

AI Technical Summary

Technical Problem

Infrared small-object detection has the problem of high false detection rates in long-distance imaging and complex backgrounds. In traditional multi-scale methods, the surrounding background grows proportionally with the growth of target cells, resulting in expansion effects and enhanced background clutter.

Method used

The multi-directional local difference measurement and entropy weighting method is adopted to calculate the local difference through sliding nested windows and multi-directional convolution kernels, and Rényi entropy weighting is introduced to design an adaptive threshold segmentation target to reduce the error detection rate.

Benefits of technology

It significantly reduces the false detection rate, enhances the target detection performance, and improves the detection ability of small infrared targets in complex backgrounds. The verification results show that they are better than other methods.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114494903B_ABST
    Figure CN114494903B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for detecting small infrared targets based on multi-directional local difference measurement and entropy weighting, comprising: step one, enhancing the target and suppressing clutter background by multi-directional local difference measurement, step two, using entropy to characterize the heterogeneity between the target and the background, further enhancing the target, and generating a saliency map at the same time. Step three, using an adaptive threshold operation to segment the target from the background. The present invention proposes a multi-directional local difference measurement method, and designs a multi-directional convolution kernel that can effectively solve the problem caused by the proportional growth of the defined background area with the growth of the target cell, and has strong robustness to the size change of the target and its surrounding local texture structure, thereby significantly reducing the false alarm rate. The present invention can effectively suppress the background while enhancing the target for infrared small target detection in a complex background, and has a higher detection rate and a low false alarm rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the field of remote sensing image target detection, and in particular to an infrared small target detection method, and in particular to an infrared small target detection method based on multi-directional local difference measurement and entropy weighting. Background Art

[0002] As one of the key technologies of infrared search and tracking systems, infrared small target detection is widely used in early warning, precision guidance and missile tracking systems. However, due to the long imaging distance, atmospheric transmission and infrared sensor errors, the target is very small and has no specific shape and texture. In addition, target detection is usually affected by edges, clutter and high-brightness complex backgrounds. These factors make infrared small target detection a challenging task.

[0003] In recent years, infrared small target detection based on the human visual system has become a research hotspot. The contrast mechanism based on the human visual system has been introduced into the detection of infrared small targets. Studies have shown that in the field of view, human vision is more sensitive to contrast than brightness. As one of the most important characteristics of human vision, the contrast mechanism is widely used in infrared small target detection methods. However, in traditional multi-scale methods based on the human visual system, the defined surrounding background grows proportionally with the growth of the target cell, which will lead to the expansion effect and the enhancement of background clutter, resulting in a higher false detection rate. Summary of the invention

[0004] Taking into account the limitations of existing methods, the present invention provides an infrared small target detection method based on multi-directional local difference measurement and entropy weighting, which not only considers the minimum directional difference between the central patch and the neighborhood background, but also pays attention to the heterogeneity of the target and the entire surrounding background, and can effectively suppress the background while enhancing the target.

[0005] The detection of small targets in infrared remote sensing images is a challenging task that requires more robust detection methods. The present invention proposes an infrared small target detection method based on multi-directional local contrast measurement and entropy weighting. The method proposes a multi-directional local difference measurement that can effectively enhance the target and suppress the background, and calculates entropy to weight the multi-directional local difference measurement to further enhance the target area. Finally, the present invention can effectively detect infrared small targets.

[0006] The specific technical solution of the present invention comprises the following steps:

[0007] Step 1: Multi-directional local difference measurement: Use a sliding nested window to slide the input image pixel by pixel from left to right and from top to bottom to enhance the local contrast of the image and obtain the local difference of a given scale; then define a multi-directional convolution kernel, convolve the convolution kernels in different directions with the input image and perform weighted operations to obtain a multi-directional local difference measurement;

[0008] The sliding nested window is used to slide the input image pixel by pixel from left to right and from top to bottom to enhance the local contrast of the image. The window is divided into a central area and a background area, where the central cell is represented by T, which is the target potential area, and Bi (i = 1, 2, ..., 8) is the background area, which is divided into 8 background cells.

[0009] The difference between the central area and the surrounding area can be defined as:

[0010] d=(d(T,B1),d(T,B2),...,d(T,B8)) T And d(T,Bi)=m T -m Bi ,(i=1,2,...,8) (1)

[0011] m T and m Bi Represent the average gray value of T and Bi respectively. In order to further measure the gray difference between the central cell and the surrounding background in different directions, we define d i , and its definition is as follows:

[0012] d i =d(T,Bi)×d(T,Bi+4),(i=1,2,…,4), (2)

[0013] where × represents a multiplication operation. Finally, the local difference at a given scale is defined as

[0014]

[0015] In order to solve the problem that the surrounding background grows proportionally with the growth of the target cell in the traditional multi-scale method, 8 convolution kernels with different directions are defined. The size of each convolution kernel at scale n is set to (n+2)×(n+2), which is composed of an n×n target window and a single pixel surrounding edge. In the target window, the pixel values ​​​​except the central pixel in a given direction are set to 1, and the pixel value at the end of the convolution kernel is set to the opposite cumulative number. Then the convolution kernels with different directions are used to perform convolution operations on the input image.

[0016] In order to solve the negative contrast caused by dark targets, we further designed the weighting coefficient, which is defined as follows:

[0017] When there are only bright targets in the scene,

[0018] When there are only dark targets in the scene,

[0019] When there are both dark and bright targets in the scene,

[0020] Among them, E(x) is the defined function, x is the input value, and finally, the image is convolved using convolution kernels in different directions and E(x) is used as the weighting coefficient;

[0021] K i =I*h i ×E(I*h i ),(i=1,2,...,8) (5)

[0022] Among them, K i is the filtering result of the original image in the i-th direction, h i is the cumulative directional filter in the i-th direction, I is the original image, and * represents the convolution operation. K i The minimum value of is defined as

[0023]

[0024] Among them, D2 is the minimum response of the directional filtering result;

[0025] Finally, the multi-directional local difference measure is defined as:

[0026] D=D1×D2 (7)

[0027] Based on the above steps, the present invention has stronger robustness to the size change of the target and its surrounding local texture structure, and significantly reduces the false detection rate.

[0028] Step 2: Entropy-based target enhancement: Use entropy-weighted multi-directional local difference measurement to obtain a saliency map;

[0029] The directional local difference measure emphasizes the minimum difference between the central patch and its directional neighborhood, but pays less attention to the heterogeneity between the central patch and the entire surrounding background. Based on the above factors, we introduce the Rényi entropy to weight the multi-directional local difference measure. The definition of the Rényi entropy is as follows:

[0030]

[0031] t is the number of possible outcomes in the variable, p j is the corresponding probability. When the Rényi entropy α=2, the entropy is more sensitive to local brightness changes, so the entropy is defined as,

[0032]

[0033] Among them, m j is the average gray value of the jth region (j = T, B, T represents the target, B represents the surrounding background). Through H2, we can measure the inconsistency between the central region and the background region. When the central region is a potential target region, its inconsistency with the background is significant, so a very small entropy value will be obtained, which can further enhance the target region.

[0034] Finally, the saliency map C is obtained through the entropy-weighted multi-directional local difference measurement, and C is defined as follows:

[0035]

[0036] In this method, the number of scales L is set to 4, and the size of the scales n = 3, 5, 7, 9. For each scale, the sliding window slides from left to right and from top to bottom, and finally the maximum value C of all scales is taken as the contrast value of the final saliency map.

[0037] Step 3: Target segmentation: For the saliency map obtained above, an adaptive threshold operation is used to segment the target from the background.

[0038] For the saliency map obtained above, the larger the contrast value, the more likely it is the target area. In order to separate the target from the background, an adaptive threshold operation is used. The threshold Th is defined as follows:

[0039] Th=μ+k×σ (11)

[0040] Among them, μ and σ represent the mean and standard deviation of the saliency map respectively, and k is an adjustable parameter. When the pixel value on the saliency map is greater than Th, it will be extracted as the target area.

[0041] The advantages and contributions of the present invention are:

[0042] 1. The present invention proposes a multi-directional local difference measurement method and designs a multi-directional convolution kernel that can effectively solve the problem caused by the proportional growth of the defined background area with the growth of the target cell. It has strong robustness to the size change of the target and its surrounding local texture structure, thereby significantly reducing the false detection rate.

[0043] 2. This method can adaptively enhance the target, and the use of entropy-weighted local difference measurement makes the present invention more robust.

[0044] 3. A large number of experiments were carried out on four real infrared image sequences with complex backgrounds, and the comparison results with current advanced methods verified the effectiveness of this method. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 Flow chart of the method of the present invention.

[0046] Figure 2 Schematic diagram of convolution kernels in different directions when the scale n=5 designed in the present invention.

[0047] Figure 3 This is a schematic diagram of different types of components of a real infrared image and their corresponding entropy values ​​in the method of the present invention.

[0048] Figure 4 These are example images from four sets of real infrared image sequences in the method of the present invention, (a) sequence 1, (b) sequence 2, (c) sequence 3, and (d) sequence 4.

[0049] Figure 5 3D grayscale distribution diagram of example images and saliency maps of different models in the method of the present invention, (a1)-(a4) 3D grayscale distribution diagram of example images, (b1)-(b4) saliency maps of model 1, (c1)-(c4) saliency maps of model 2, (d1)-(d4) saliency maps of model 3, (e1)-(e4) saliency maps of model 4, (f1)-(f4) saliency maps of model 5, (g1)-(g4) saliency maps of model 6, (h1)-(h4) saliency maps of model 7. DETAILED DESCRIPTION

[0050] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0051] In order to better understand the technical route of the present invention, the specific implementation of the present invention is described below with reference to the accompanying drawings.

[0052] The present invention is a method for detecting small infrared targets by using directional local difference measurement and entropy weighting. The flow chart is as follows: Figure 1 As shown, the specific steps include:

[0053] Step 1: Multi-directional local difference measurement

[0054] The input real infrared image I is filtered using a nested window to calculate the difference between the target and the surrounding background, which can suppress the background while enhancing the target area. In order to solve the problem of the background area growing proportionally with the target cell, we define a multi-directional convolution kernel, and convolve 8 convolution kernels in different directions with the input image at a given scale.

[0055] The sliding nested window is divided into a central area and a background area, where the central cell is represented by T, which is the target potential area, and B i (i=1,2,…,8) is the background area, and the background area is divided into 8 background cells.

[0056] The difference between the central area and the surrounding area can be defined as:

[0057] d=(d(T,B1),d(T,B2),...,d(T,B8)) T And d(T,Bi)=m T -m Bi ,(i=1,2,…,8) (1)

[0058] m T and m Bi Represent the average grayscale values ​​of T and Bi respectively. In order to further measure the grayscale difference between the central area and the background area in different directions, we define d i , and its definition is as follows:

[0059] d i =d(T,Bi)×d(T,Bi+4),(i=1,2,...,4), (2)

[0060] where × represents the multiplication operation. Finally, the local difference at a given scale is defined as:

[0061]

[0062] Define 8 convolution kernels in different directions, each convolution kernel is set to (n+2)×(n+2) in scale n, and is composed of an n×n target window and a surrounding edge of a single pixel. In the target window, the pixel values ​​​​except the central pixel in a given direction are set to 1, and the pixel value at the end of the convolution kernel is set to the opposite cumulative number. In the present invention, the number of scales L is set to 4, and the scales n=3, 5, 7, 9. Figure 2 When the scale n=5, there are 8 convolution kernels in different directions.

[0063] Secondly, in order to solve the negative contrast caused by dark targets, we further designed the weighting coefficient, which is defined as follows:

[0064] When there are only bright targets in the scene,

[0065] When there are only dark targets in the scene,

[0066] When there are both dark and bright targets in the scene,

[0067] Among them, E(x) is the defined function, x is the input value, and finally, the image is convolved using convolution kernels in different directions and E(x) is used as the weighting coefficient to obtain the following expression;

[0068] K i=I*h i ×E(I*h i ),(i=1,2,…,8)(5)

[0069] Among them, K i is the filtering result of the original image in the i-th direction, h i is the cumulative directional filter in the i-th direction, I is the original image, * represents the convolution operation, K i The minimum value of is defined as:

[0070]

[0071] Among them, D2 is the minimum response of the directional filtering result;

[0072] Finally, the multi-directional local difference measure is defined as:

[0073] D=D1×D2 (7)

[0074] Step 2: Entropy-based target enhancement

[0075] Using a sliding nested window, the original image is filtered, the grayscale mean of the central cell and the surrounding background is calculated, and then the Rényi entropy is introduced to weight the multi-directional local difference measure. The definition of the Rényi entropy is as follows:

[0076]

[0077] t is the number of possible outcomes in the variable, p j is the corresponding probability. When the Rényi entropy α=2, the entropy is more sensitive to local brightness changes, so the entropy is defined as,

[0078]

[0079] Among them, m j It is the average gray value of the jth (j=T, B, T represents the target and B represents the surrounding background) region.

[0080] Figure 3 The different types of components in the real infrared image and their corresponding entropy values. When the sliding window slides to the target area and its central cell is exactly the size of the target area, the background and the target area have significant inconsistency, so a very small entropy value will be obtained, which can further enhance the target area. Conversely, when the sliding window slides to the background area, a larger entropy value will be obtained, which plays a role in suppressing the background.

[0081] Finally, the saliency map C is obtained through the entropy-weighted multi-directional local difference measurement, and C is defined as follows:

[0082]

[0083] In this method, the number of scales L is set to 4, and the size of each scale n = 3, 5, 7, 9. For each scale, the sliding window slides from left to right and from top to bottom, and finally the maximum value C at different scales is taken as the contrast value of the final saliency map.

[0084] Step 3: Target segmentation

[0085] For the saliency map obtained above, the larger the contrast value, the more likely it is the target area. In order to segment the target from the background, an adaptive threshold operation is used, and the threshold Th is defined as follows:

[0086] Th=μ+k×σ (11)

[0087] Among them, μ and σ represent the mean and standard deviation of the saliency map respectively, and k is an adjustable parameter. When the pixel value on the saliency map is greater than Th, it will be extracted as the target area.

[0088] In order to verify the superiority of this method, we selected 4 groups of representative real infrared image samples (such as Figure 2 As shown in Figure 2, the experimental results are compared with 6 infrared small target detection models. Figure 4 The three-dimensional grayscale distribution of image samples and enhanced images of different models are shown. The steps are as follows:

[0089] (1) Four representative real infrared image sequences were selected

[0090] (2) Six infrared small target models are selected, namely:

[0091] Model 1: [CLPChen, H.Li, Y.Wei, T.Xia, and YYTang, "A local contrast method for small infrared target detection," IEEE Trans.Geosci.Remote Sens., vol.52, no.1, pp.574–581, Jan.2014.]

[0092] Model 2: [Y.Wei,

[0093] Model 3: [J.Han, K.Liang, B.Zhou,

[0094] Model 4: [S.Aghaziyarati, S.Moradi, and H.Talebi, "Small infrared target detection using absolute average difference weighted by cumulative directional derivatives," Infr.Phys.Technol., vol.101, pp.78–87, Sep.2019.]

[0095] Model 5: [P.Du and A.Hamdulla, "Infrared small target detection using homogeneity-weighted local contrast measure," IEEE Geosci.Remote Sens.Lett., vol.17, no.3, pp.514–518, Mar.2020.]

[0096] Model 6: [X.Kong, C.Yang, S.Cao, C.Li, and Z.Peng, "Infrared Small TargetDetection via Nonconvex Tensor Fibered Rank Approximation," IEEETrans.Geosci.Remote Sens., pp.

[0097] 1-21.2021.]

[0098] Model 7: Infrared small target detection method based on multi-directional local difference measurement and entropy weighting

[0099] (2) Experiments were conducted on seven models and compared under four sets of image sequences, and the characteristic indicators selected were: indicator 1: background suppression factor, indicator 2: signal-to-noise ratio gain.

[0100] Table 1

[0101]

[0102] (3) The results in Table 1 show that the present invention has better detection performance than other models. Figure 5 In the salient map of the example image shown, the salient map of the present invention effectively suppresses the background while enhancing the target, which is superior to other models. Therefore, the present invention can effectively and robustly detect small infrared targets with complex backgrounds.

[0103] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

Claims

1. A method for detecting small infrared targets based on multi-directional local difference measurement and entropy weighting, characterized in that: The steps include: Step 1: Use a sliding nested window to slide the input image pixel by pixel from left to right and from top to bottom to enhance the local contrast of the image and obtain the local difference of a given scale; then define a multi-directional convolution kernel, convolve the convolution kernels in different directions with the input image and perform weighted operations to obtain a multi-directional local difference measure; The nested window slid in step 1 is divided into a central area and a background area, where the central cell is represented by T, which is the target potential area, and Bi is the background area, which is further divided into 8 background cells, i = 1, 2, ..., 8; The difference between the central area and the surrounding area can be defined as: d = (d(T, B1), d(T, B2),..., d(T, B8)) T and d(T, Bi) = m T -m Bi , (i = 1, 2,..., 8) (1) Among them, m T and m Bi Represent the average grayscale values ​​of T and Bi respectively. In order to further measure the grayscale difference between the central area and the background area in different directions, define d i as follows: d i =d(T,Bi)×d(T,Bi+4),(i=1,2,...,4), (2) where × represents the multiplication operation. Finally, the local difference at a given scale is defined as: Eight convolution kernels in different directions are defined. The size of each convolution kernel in scale n is set to (n+2)×(n+2), which is composed of an n×n target window and the surrounding edge of a single pixel. In the target window, the pixel values ​​​​except the central pixel in a given direction are set to 1, and the pixel values ​​​​at the end of the convolution kernel are set to the opposite cumulative number; secondly, in order to solve the negative contrast caused by dark targets, the weighting coefficient is further designed, which is defined as follows: When there are only bright targets in the scene, When there are only dark targets in the scene, When there are both dark and bright targets in the scene, Among them, E(x) is the defined function, x is the input value, and finally, the image is convolved using convolution kernels in different directions and E(x) is used as the weighting coefficient; K i =I*h i ×E(I*h i ),(i=1,2,…,8) (5) Among them, K i is the filtering result of the original image in the i-th direction, h i is the cumulative directional filter in the i-th direction, I is the original image, * represents the convolution operation, K i The minimum value of is defined as: Among them, D2 is the minimum response of the directional filtering result; Finally, the multi-directional local difference measure is defined as: D = D1 × D2 (7); Step 2: Use entropy-weighted multi-directional local difference measurement to obtain a saliency map; Step 3: For the saliency map obtained above, an adaptive threshold operation is used to segment the target from the background.

2. The infrared small target detection method based on multi-directional local difference measurement and entropy weighting as claimed in claim 1, characterized in that: The definition of entropy in step 2 is as follows; Apply Rényi entropy to weight the local difference measure D, which is defined as: t is the number of possible outcomes in the variable, p j is the corresponding probability. When the Rényi entropy α = 2, the entropy is more sensitive to local brightness changes, so the entropy is defined as: Among them, m j is the average gray value of the jth region, j = T, B, T represents the target, B represents the surrounding background, and the inconsistency between the central region and the background region is measured by H2. When the central region is a potential target region, its inconsistency with the background is significant, so a very small entropy value will be obtained, which can further enhance the target region; Finally, the saliency map C is obtained through the entropy-weighted multi-directional local difference measurement, and C is defined as follows:

3. The infrared small target detection method based on multi-directional local difference measurement and entropy weighting as claimed in claim 1, characterized in that: The threshold Th in step 3 is defined as follows: Th=μ+k×σ (11) Among them, μ and σ represent the mean and standard deviation of the saliency map respectively, and k is an adjustable parameter. When the pixel value on the saliency map is greater than Th, it will be extracted as the target area.

4. The infrared small target detection method based on multi-directional local difference measurement and entropy weighting as claimed in claim 1, characterized in that: Scale n=3,5,7,9.