Infrared small target detection method for improving multi-scale local contrast

Abstract

The invention discloses an infrared small target detection method for improving multi-scale local contrast, and the method comprises the steps: constructing diffusion coefficient functions through employing gray difference product values in four different directions in a local region, taking a minimum function value in the four diffusion coefficient functions as a final filtering result, and multiplying the final filtering result with an original image to obtain a final detection effect image; according to the method, the local contrast is combined with the anisotropic diffusion coefficient, so that the target is reserved while the complex background is effectively inhibited.

Description

Technical Field

[0001] The invention belongs to the technical field of infrared detection, and in particular relates to an infrared small target detection method with improved multi-scale local contrast. Background Art

[0002] Infrared detection systems have been widely used in guidance, early warning and other fields due to their strong anti-interference capabilities. However, in actual IRST systems, targets at long distances present small target characteristics, that is, they occupy a very small proportion in the image and have no texture structure information. Moreover, in complex backgrounds and low signal-to-noise ratio conditions, small targets are easily obliterated by background clutter and noise.

[0003] In recent years, the robust human visual system (HVS) characteristics have been introduced into the field of infrared dim target detection. According to the contrast mechanism of HVS, the most important part in our visual system is contrast, not brightness. This fact is of great value for infrared dim target detection. Early HVS algorithms calculated the contrast by the grayscale ratio of the target and the surrounding local background, such as the local contrast measurement algorithm. They used the grayscale ratio of the center and the neighborhood as the enhancement coefficient of the center point, which can effectively enhance the target, but the ability to suppress the highlight background is limited, and the "diffusion effect" of the detected target is serious. After that, many scholars have made many improvements to this type of algorithm. Wei et al. proposed a contrast measurement based on multi-scale blocks, which uses the grayscale difference between the target and the symmetric single neighborhood to achieve rapid detection of bright and dark targets, but the suppression effect on severe clutter is weak, and the enhancement effect of weak targets in multiple targets is not good. Wu et al. proposed a dual-neighborhood gradient method to adapt to multi-scale targets. DNGM uses dual-neighborhood grayscale difference to enhance the target, which makes the enhancement effect of weak targets much smaller than that of strong targets, resulting in poor detection of weak targets in multiple targets. Lv et al. proposed a regional intensity level (RIL) algorithm as a weighted function of local contrast, but due to the local regional maximum value used in RIL, it is easily disturbed by high brightness pixel size noise (PNHB), the background suppression effect is not thorough, and the weak target enhancement effect is poor. Therefore, Han et al. proposed an improved RIL (IRIL) algorithm to replace the maximum value with the average value calculated by multiple maximum grayscales, and Cui et al. proposed a new regional intensity level (NRIL) algorithm to weight the local contrast map to further suppress the background, but such algorithms cannot enhance multiple targets at the same time, and the weak target detection effect in multiple targets is poor.

[0004] In summary, there is currently no algorithm that can enhance multiple targets simultaneously. Summary of the invention

[0005] In view of this, the present invention provides an infrared small target detection method with improved multi-scale local contrast, which can effectively suppress complex background while retaining the target.

[0006] The technical solution for implementing the present invention is as follows:

[0007] An infrared small target detection method with improved multi-scale local contrast comprises the following steps:

[0008] Step 1: Scan the entire image with a U-shaped window, with the inner window as the center area and the outer window as the background area;

[0009] Step 2: Calculate the mean gray value m of the central area T and the mean gray value of the background area m B ; In a U-shaped window, with the central area T as the center, the background area is divided into 8 sub-areas: upper left, upper, upper right, left, right, lower left, lower, and lower right, denoted as B z , z=1,2,...,8, calculate the local grayscale contrast d(T,B z );

[0010] Step 3: Calculate the local contrast value D k , and according to D k Calculate the anisotropic diffusion coefficient to obtain the diffusion values ​​C in four directions k ;

[0011] Step 4: Take the smallest diffusion value in the four directions as the output value, and take the smallest output value in all windows of the entire image as the final output value W; multiply the final output value W by the original image to obtain the final detection effect image.

[0012] Furthermore, the U-shaped window includes three pixel scales, namely, 3*3, 5*5, and 7*7 pixels. During the scanning process, each scale of the window traverses and scans the entire image.

[0013] Further, Represents the average gray value of the z-th background sub-region.

[0014] Further,

[0015] Further, Where A is a constant, which is 100. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the U-shaped window model of the present invention.

[0017] Figure 2 The present invention targets situations in different regions.

[0018] Figure 3The detection results of various algorithms for pictures in the embodiment; the sequence image S1 represents an infrared image with fixed pattern noise, in which non-uniform noise with vertical stripes as the main feature is prone to false alarm points;

[0019] Sequence image S2 shows a cloudy sky background. Cloud edges with dramatic grayscale fluctuations and irregular block clouds are common clutter in sky scenes.

[0020] The sequence image S3 represents multiple targets with different brightness under the sky background. DETAILED DESCRIPTION

[0021] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0022] The present invention provides a multi-scale local contrast infrared small target detection method (ADMPCM) based on anisotropic diffusion coefficient, which constructs a diffusion coefficient function by utilizing the grayscale difference product value in four different directions in a local area, and takes the minimum function value among the four diffusion coefficient functions as the final filtering result.

[0023] The specific process of the method of the present invention is as follows:

[0024] Small targets are prominent in local areas but not in the entire image, so the present invention mainly focuses on small local image blocks (such as Figure 1 As shown in the figure, the pixel value of the central pixel of the target area is defined by using the difference calculation between the target area and the background area. The central area is used to capture the real target, and the peripheral area is used to capture the surrounding background. To ensure that the cell size N is close to or slightly larger than the real target, the window model N here selects three pixel templates of 3*3, 5*5, and 7*7 to ensure that targets of different sizes are detected while introducing as little interference as possible.

[0025] Figure 1 The nested structure window model shown in , where the central area is marked with a red box to represent the target, and the surrounding background area is marked with a white box. The sliding window moves from left to right and from top to bottom pixel by pixel in the infrared image, as shown in Figure 1 As shown. Central area m T and background area m B The average gray value is:

[0026]

[0027] Where N T is the number of pixels in the target area, υ i Represents the grayscale value of the i-th pixel in the target area. Similarly, N B is the number of pixels in its corresponding background area, υ jis the gray value of the jth pixel in the background area.

[0028] Then, the average value in the middle window is compared with the adjacent windows to calculate the local grayscale contrast d(T,B z ),

[0029]

[0030] Where m T and are the average grayscale values ​​of the middle window and the zth background window respectively.

[0031] According to formula 3, the local grayscale contrast values ​​in the corresponding four directions are multiplied to obtain the local contrast value D k . It can be expressed as,

[0032]

[0033] After obtaining the local contrast value D k After that, the diffusion values ​​C in four directions are obtained through the anisotropic diffusion coefficient. k , the specific formula is as follows:

[0034]

[0035] Where A is a constant, which is 100.

[0036] Then, take the smallest diffusion value in the four directions as the output value W N :

[0037] W N =min(C k ) (5)

[0038] Finally, the minimum diffusion value under different templates is taken as the final output value W:

[0039] W=min(W N ) (6)

[0040] The final output value W is multiplied by the original image to obtain the final detection effect map. The final infrared small target detection map ADMPCM can be defined as:

[0041] ADMPCM(x,y)=W(x,y)*υ(x,y) (7)

[0042] Among them, υ(x,y) represents the grayscale value of the pixel with image coordinates (x,y).

[0043] Effect analysis

[0044] like Figure 2As shown in the figure, TT indicates that the central area is the real target, NB indicates that the central area is the normal sky background pixel, HB indicates that the central area is the high-brightness background pixel, EB indicates that the central area is the complex background edge pixel, and PNHB indicates that the central area is random noise.

[0045] (1) For the NB and HB regions, when the sliding window passes through the NB and HB regions, the local contrast values ​​in the four directions are close to 0, and the diffusion coefficient values ​​in the four directions are all close to 0. In the final output ADMPCM image, the grayscale values ​​of these two regions are suppressed to the greatest extent.

[0046] (2) For the EB region, when the sliding window passes through the EB region, in the direction where the cloud changes slowly, the local contrast value is close to 0 and the diffusion coefficient value is close to 0. In the direction where the cloud changes quickly, the local contrast value is negative and is set to zero, and the diffusion coefficient value is 0. Therefore, false alarms that occur at the edge of the cloud can be effectively suppressed.

[0047] (3) For the PNHB area, when the sliding window passes through the PNHB area, the grayscale value of the random noise is close to the true target grayscale, but the random noise usually occupies a very small number of pixels. When the algorithm calculates the average, the random noise grayscale value is dispersed, making the local contrast value smaller. From the diffusion function, it can be seen that when the contrast value is small, the diffusion function value changes extremely slowly, and the diffusion coefficient value is close to 0. Therefore, the influence of random noise can be effectively eliminated by the diffusion coefficient.

[0048] (4) For the TT area, when the sliding window passes through the TT area and the center of the window covers the target in the background, a positive difference is generated between the central window and the outer window, and the difference is further amplified by the product to obtain a larger local contrast value. After the diffusion coefficient operation, a value very close to 1 is obtained. Only in the target area will the diffusion function values ​​in the four directions be close to 1. Therefore, the background is greatly suppressed by the diffusion function, the target gray value is retained, and the "diffusion effect" brought by the traditional local contrast method is eliminated.

[0049] Example

[0050] Figure 3 The first column shows three real infrared image sequences used in the present invention. Sequence image S1 represents an infrared image with fixed pattern noise, in which non-uniform noise with vertical stripes as the main feature easily produces false alarm points. Sequence image S2 represents a cloudy sky background. Cloud edges with drastic grayscale fluctuations and irregular block clouds are common clutter in sky scenes. Sequence image S3 represents multiple targets under the sky background. Multiple targets with different brightness are the main cause of missed detection. The above sequences represent different scenes, respectively, and contain multiple types of clutter.

[0051] The proposed method (ADMPCM) is compared with five traditional HVS-based methods, including DLCM, ADMD, HBLCM, RLCM, and MPCM. The DLCM algorithm is a diagonal grayscale difference contrast measurement method based on a double-layer sliding window, RLCM is a multi-scale relative local contrast measurement method, ADMD is a target detection method using absolute directional average difference, HBLCM is an algorithm that uses high-boost to sharpen images to enhance targets, and MPCM refers to a block-based contrast measurement method.

[0052] The target enhancement performance is mainly based on the single-frame SCR gain, and the SCR is calculated as follows:

[0053]

[0054] Where T average is the average brightness of the target, μ is the average brightness of the background, σ is the standard deviation of the image, and SCRG is defined as

[0055]

[0056] Among them, SCR out is the image signal-to-noise ratio after the ADMPCM algorithm, SCR in is the image signal-to-noise ratio of the original image. Accordingly, the performance evaluation of background suppression is mainly based on the background suppression factor, which is

[0057]

[0058] Among them, σ in is the standard deviation of the original image, σ out is the standard deviation after the algorithm.

[0059] Table 1 SCRG of each algorithm

[0060]

[0061] Table 2 BSF of each algorithm

[0062]

[0063] Figure 3The comparison of experimental results of some representative images of sequence images S1, S2, and S3 is shown. As can be seen from S1, the gray value of the target in the sky is close to the infrared image stripe noise, which increases the difficulty of detection. There are many false detections in ADMD, RLCM, HBLCM, and MPCM. Only DLCM and ADMPCM can accurately extract the target. When the contrast between the target and the background is very low, as shown in the S2 sequence, all methods can accurately extract the target, but the "extension effect" of the RLCM method is obvious, resulting in the change of the target shape in the saliency map. ADMPCM successfully overcomes the "extension effect" and can accurately detect the number and position of small targets. The sky in the original image of the S3 sequence contains six small moving targets. Due to the large differences in the grayscale and size of each target, except for ADMPCM, which can effectively enhance the six small infrared targets at the same time, other algorithms cannot effectively enhance the six small infrared targets, and the grayscale value gap between weak targets and strong targets is widened, which will greatly affect the accuracy of target detection.

[0064] Tables 1 and 2 give the SCR and BSF values ​​of each method on the test image sequences. In all sequences, ADMPCM obtained the maximum SCRG value. In single target detection, the SCRG value increased by up to twice, and in multi-target detection, the SCRG value of weak targets increased by up to 10 times. For the BSF value, ADMPCM reached the highest value on all datasets, with an average increase of 2 times. This means that ADMPCM has better target enhancement and clutter suppression effects than other methods. In summary, the ADMPCM algorithm can obtain more accurate detection results when detecting small infrared targets.

[0065] Beneficial effects:

[0066] 1. The present invention combines local contrast with anisotropic diffusion coefficient to effectively suppress complex background while retaining the target.

[0067] 2. The present invention can effectively detect small infrared targets of different scales through multi-scale templates.

[0068] 3. The SCRG value and BSF value of the common target detection indicators of the method of the present invention have reached the highest value in the data set, and have increased by 2 times on average. This means that the method of the present invention has better target enhancement and clutter suppression effects than other methods.

[0069] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An infrared small target detection method with improved multi-scale local contrast, characterized in that: The following steps are involved: Step 1: Scan the entire image with a U-shaped window, with the inner window as the center area and the outer window as the background area; Step 2: Calculate the mean gray value m of the central area T and the mean gray value of the background area m B ; In a U-shaped window, with the central area T as the center, the background area is divided into 8 sub-areas: upper left, upper, upper right, left, right, lower left, lower, and lower right, denoted as B z , z=1,2,...,8, calculate the local grayscale contrast d(T,B z ); Step 3: Calculate the local contrast value D k , and according to D k Calculate the anisotropic diffusion coefficient to obtain the diffusion values ​​C in four directions k ; Step 4: Take the smallest diffusion value in the four directions as the output value, and take the smallest output value in all windows of the entire image as the final output value W; Multiply the final output value W by the original image to obtain the final detection effect image.

2. The infrared small target detection method according to claim 1, characterized in that: The U-shaped window includes three pixel scales, namely, 3*3, 5*5, and 7*7 pixels. During the scanning process, each scale of the window traverses and scans the entire image.

3. The infrared small target detection method according to claim 1, characterized in that: Represents the average gray value of the z-th background sub-region.

4. The infrared small target detection method according to claim 3, characterized in that:

5. The infrared small target detection method according to claim 1 or 4, characterized in that: Where A is a constant, which is 100.

Images