CN113516187A

CN113516187A - Infrared weak and small target detection algorithm adopting local characteristic contrast

Info

Publication number: CN113516187A
Application number: CN202110790994.1A
Authority: CN
Inventors: 韩金辉; 陈园园; 张鸿辉; 朱欣颖; 李娜娜; 姚遥; 陈耀弘; 赵骞; 李知铮; 桑晓丹; 赵劼; 袁旭野
Original assignee: Zhoukou Normal University
Current assignee: Zhoukou Normal University
Priority date: 2021-07-13
Filing date: 2021-07-13
Publication date: 2021-10-19

Abstract

The invention provides an infrared dim target detection algorithm adopting local characteristic contrast, which belongs to the field of image processing and comprises the following steps: acquiring an original image, and expanding the existing three-layer window to obtain a new nested window consisting of an inner layer and eight outer layers; respectively calculating the local contrast of the inner layer and the outer layer of the new nested window by adopting a ratio difference combination method, taking the local contrast as the characteristics of the inner layer and the outer layer, and calculating the ratio difference combination type contrast between the inner layer and the outer layer to obtain the characteristic contrast; designing a weighting function by utilizing the isolation layer of the new nested window, and weighting the characteristic contrast so as to further inhibit a complex background and highlight a target; and extracting the infrared dim targets in the weighted image by using threshold operation. Compared with the existing local contrast algorithm, the algorithm can obtain better target enhancement capability and background suppression capability, and has certain advantages in the aspects of detection rate, false alarm rate, instantaneity and the like.

Description

Infrared weak and small target detection algorithm adopting local characteristic contrast

Technical Field

The invention belongs to the field of image processing, and particularly relates to an infrared dim target detection algorithm adopting local feature contrast.

Background

Infrared detection systems have been widely used in many fields such as guidance and early warning. However, when the detection distance is very far, the area occupied by the target in the original image output by the system is small and the brightness is generally weak under the influence of factors such as the parameter limit of the optical system and atmospheric attenuation, and the target is called as an infrared weak target.

The method has very important theoretical significance and practical value for effectively detecting the infrared dim target, but always faces great difficulty, and the main reasons include: firstly, the target has small area and weak brightness, lacks sufficient and obvious information such as color, shape, texture and the like, is difficult to directly detect, and causes the detection rate of the target to be lower; secondly, in typical application scenes such as sea, sky and the like, a background with extremely high brightness and very complex edges may appear in a detector view field, so that detection is easily interfered, and more false alarms are caused; meanwhile, factors such as defects of the production process of the focal plane and electrical noise during the operation of the device can cause a series of noise interferences and can also cause false alarms. Then, in many practical applications, the scale of the target is usually unknown information, so that the selection of algorithm parameters becomes very difficult; finally, some practical applications have a high requirement on real-time detection, which brings further difficulty to the detection of small targets.

In recent years, researchers at home and abroad propose a plurality of infrared weak and small target detection algorithms which can be roughly divided into two categories, namely a sequence type and a single-frame type. The sequence detection algorithm directly extracts targets among a plurality of adjacent frames, utilizes more information, and has better detection performance generally, but the data volume and the calculated amount of the sequence detection algorithm are larger, the output of the sequence detection algorithm often has certain hysteresis, and the requirement of some practical applications on real-time performance is difficult to meet. The single-frame type detection algorithm detects the target in one frame, the data amount and the calculated amount are smaller than those of the sequence type algorithm, the real-time performance is generally better, and meanwhile, some single-frame type algorithms can also be used as basic modules of some sequence type algorithms.

Many types of single frame type detection algorithms have been proposed, such as background estimation based, morphology based, directional derivative or gradient based, frequency domain filtering based, sparse representation based, sparse low rank decomposition based, local contrast mechanism based, etc. The algorithm based on the local contrast mechanism is used as one of the bionic algorithms, and is mainly inspired by the contrast mechanism of the human visual system, namely, human eyes are more sensitive to contrast information in a visual field than brightness information. In the original ir image, the object, although often not the brightest part of the entire image, is usually slightly brighter than its surrounding background (essentially because typical valuable objects, such as airplanes, vehicles, ships, etc., are generally hotter than their surrounding background). By extracting the local contrast information in the image and taking the local contrast information as the basis of target detection, the influence of a complex background can be effectively eliminated, and better detection performance is obtained. Therefore, detection algorithms based on local contrast mechanisms have been particularly emphasized in recent years.

The nature of local contrast is the information of the difference between the current position and its surrounding neighbourhood, but is not yet publicly known in terms of specific definition. Different researchers have proposed many methods for calculating local contrast according to their needs. According to the calculation method of the contrast, it can be classified into three basic types of a gray scale difference type, a gray scale ratio type, and a gray scale ratio difference combination type. Meanwhile, the DLCM (Double-layer Local Contrast Measure) algorithm proposed by pannaga et al is one of the important bases of the algorithm herein, and its main contribution is to propose a novel three-layer window, as shown in fig. 1. The window contains 5 × 5 small sub-blocks, and each small sub-block is 3 × 3 pixels in size. The central sub-block T is used for capturing main energy near the center of the target; the peripheral sub-blocks SB 1-SB 16 are used for capturing the neighborhood background of the target and carrying out contrast calculation with the target; the intermediate sub-blocks IB 1-IB 8 are used for isolating the target from the background, so that the algorithm can directly solve the detection of small targets with unknown sizes by using single-scale calculation without adjusting the window size.

In DLCM, the minimum value of the gray difference between the central sub-block and the peripheral sub-blocks is taken as the contrast information D of the current position, that is, the contrast information D is

D＝min_j d(T,SB_j),j＝1,2,...,16 (2)

In the formula, m₀And m_SBjRespectively representing the central subblock T and the surrounding subblocks SB_jThe average value j is 1 to 16, which are the numbers of the surrounding sub-blocks. Note that non-negative constraints are used in equation (1) to eliminate clutter residues.

Meanwhile, the DLCM also introduces a weighting operation using information of the intermediate layer (i.e., the isolation layer), and defines a weighting function as:

W＝min_i[d(T,IB_i)×d(T,IB_9-i)],i＝1,2,3,4 (4)

in the formula, m₀And m_IBiRespectively representing the central subblock T and the intermediate subblock IB_iThe average value of the gray levels of (1) to (8) is the serial number of the intermediate layer sub-block. The non-negative constraint is also used in equation (3).

The final contrast information for the current position would then be:

DLCM＝W×D (5)

the window is used for traversing the original image, and then the DLCM calculation result of each pixel point can be obtained. The authors think that the DLCM value of the real target will be the most prominent, so the target can be extracted using threshold segmentation.

Theoretically, at present, the performance of the proportional-difference combined contrast algorithm is the best among the infrared weak and small target detection algorithms based on the local contrast mechanism, and researchers have tried many improving means, however, the current improvement still has certain disadvantages, including:

in the prior art, the various windows used by the algorithm only focus on the difference between the central region and the surrounding neighborhood background, but the characteristics of the neighborhood background are not considered sufficiently.

Second, some researchers have proposed that contrast can be obtained by using features other than gray scale, but the features currently selected by people are simpler, such as variance, gradient, etc. When the background is complex, it is difficult for these features to accurately describe local information of the image.

Weighting the original contrast information by using a weighting function is an effective means for improving the detection performance, but some weighting functions proposed by people at present are usually complex in calculation, so that the overall algorithm structure is complex.

Aiming at the problems, the invention provides an infrared weak and small target detection algorithm adopting local characteristic contrast.

Disclosure of Invention

In order to overcome the defects of the prior art, the invention provides an infrared weak and small target detection algorithm adopting local characteristic contrast.

In order to achieve the above purpose, the invention provides the following technical scheme:

an infrared weak and small target detection algorithm adopting local characteristic contrast comprises the following steps:

step 1, acquiring an original image, and expanding the existing three-layer window to obtain a new nested window consisting of an inner layer and eight outer layers, wherein the inner layer window is used for inspecting the data characteristics of the current position, and the outer layer window is used for inspecting the data characteristics of the local neighborhood of the current position;

step 2, respectively calculating the local contrast of the inner layer and the outer layer of the new nested window by adopting a ratio difference combination method, taking the local contrast as the characteristics of the inner layer and the outer layer, and calculating the ratio difference combination type contrast between the inner layer and the outer layer to obtain the characteristic contrast;

step 3, designing a weighting function by utilizing the isolation layer of the new nested window, and weighting the characteristic contrast to further inhibit a complex background and highlight a target;

and 4, extracting the infrared dim targets in the weighted image by using threshold operation.

Preferably, the new nested window is composed of 9 × 9 sub-blocks in total, and each sub-block is 3 × 3 pixel points in size; wherein, the middle 5 × 5 sub-blocks form an inner layer window, and comprise a central sub-block T and peripheral sub-blocks SB₁～SB₁₆And intermediate subblock IB₁～IB₈Composition is carried out;

outside the nested windows, respectively with L₁₀，L₂₀，…，L₈₀For the center, every 3 × 3 subblocks constitutes an outer window.

Preferably, the step 2 comprises the following steps:

step 2.1, extraction of inner layer characteristics

Step 2.1.1, Gaussian filtering is carried out on the central subblock

A 3 x 3 two-dimensional Gaussian filter template similar to the infrared dim target in shape is adopted to filter the central subblock T, so that the signal-to-noise ratio of an original image is improved;

at the pixel point (i, j), the matched filtering result is defined as:

wherein T is a central subblock, P and q are intermediate variables, G is a filtering template, and IG_inIs the result after matching the filtering;

step 2.1.2, nearest filtering is carried out on surrounding sub-blocks

The inner window includes SB₁～SB₁₆And in total, 16 surrounding sub-blocks, wherein the surrounding sub-block closest to the Gaussian filtered value of the central sub-block participates in calculation:

in the formula, (i, j) is the coordinate of the current pixel point, IG_inGaussian filtered value, m, representing the center subblock_SBnRepresents the mean of the nth surrounding sub-block, B_inRepresenting the final background value of the selected inner layer window;

step 2.1.3, calculating the contrast of the inner window by adopting the ratio difference combination

And (3) adopting a calculation mode of ratio-difference combination to enhance a real target and inhibit a complex background:

in the formula, RD_inThe ratio difference joint contrast of the inner layer window is represented, and xi is a constraint factor;

step 2.2, extraction of outer layer characteristics

Outer layer window containment L for a newly nested window₁～L₈For the nth outer window, the center sub-block is first gaussian filtered:

in the formula L_n0Is the central sub-block of the nth outer window, G is the filtering template, IG_outnIs the result after matching the filtering;

then, selecting the peripheral sub-blocks closest to the Gaussian filter value of the central sub-block to participate in calculation;

in the formula, L_niIs the i-th surrounding sub-block of the n-th outer window, m_LniMeans, IG, of the i-th surrounding sub-block of the n-th outer window_outnRepresenting the Gaussian filtered value of the central sub-block of the nth outer window, B_outnRepresenting the final background value of the nth skin window;

the specific difference of the nth skin window then gives the combined contrast:

step 2.3, calculation of feature contrast

Firstly, according to the closest principle, selecting the value which is closest to the characteristics of the inner layer window from 8 outer layer windows to participate in the operation, namely:

then, the characteristic contrast between the inner window and the outer window is calculated by using a ratio difference joint mode:

preferably, in step 3, only when the maximum pixel value of the center sub-block of the inner layer window is greater than the average pixel value of any intermediate sub-block, the weight W (i, j) at the position is made to be 1, and otherwise, the weight W (i, j) is made to be 0;

the final saliency map SM will be the product of the feature contrast and the weighting function:

SM(i,j)＝RD(i,j)gW(i,j) (14)

where W is the weighting function, RD is the characteristic contrast, and SM is the final result after weighting.

Preferably, the step 4 comprises the following steps:

firstly, normalizing SM to a range of 0-1, and then defining a threshold TH as follows:

TH＝μ+k·std (15)

in the formula, mu is the mean value of SM; std is the standard deviation of SM; k is a coefficient;

using TH to carry out binarization operation on SM, marking the pixel points larger than TH as 1, and marking the pixel points smaller than TH as 0; and performing expansion operation on the marking result, and finally outputting each connected region with the value of 1 as a detected target region.

The infrared dim target detection algorithm adopting the local characteristic contrast provided by the invention has the following beneficial effects:

(1) the invention designs a novel nested window, which consists of an inner layer window and eight outer layer windows, wherein the inner layer window can be used for inspecting the data characteristics of the current position, and the outer layer window can be used for inspecting the data characteristics of the local neighborhood of the current position;

(2) the local contrast of the inner window and the outer window is selected as a feature, and the feature contrast is calculated between the inner window and the outer window, so that compared with the existing algorithms which directly use the gray value to calculate the contrast, the method can better describe the local information of the image when the background is complex, and the calculation result is more accurate;

(3) by utilizing the isolation layer, a weighting function which is simple and convenient to calculate but very effective is designed, so that the final detection performance is effectively improved, and certain advantages are achieved in the aspects of calculation complexity and detection real-time performance.

Drawings

In order to more clearly illustrate the embodiments of the present invention and the design thereof, the drawings required for the embodiments will be briefly described below. The drawings in the following description are only some embodiments of the invention and it will be clear to a person skilled in the art that other drawings can be derived from them without inventive effort.

FIG. 1 is a three-layer window structure used by the existing DLCM algorithm;

FIG. 2 is a flowchart of an infrared small and weak target detection algorithm using local feature contrast according to embodiment 1 of the present invention;

FIG. 3 is a nested window designed by the algorithm of the present invention;

FIG. 4 is a Gaussian filter template;

FIG. 5 is a schematic of the calculations at different types of pixel points;

FIG. 6 is a schematic diagram of the detection process and results of different sequences using the algorithm of the present invention;

FIG. 7 is a schematic diagram of the detection process and results of 6 single-frame images using the algorithm of the present invention;

FIG. 8 is a graph comparing ROC curves for 8 sequences for different algorithms;

FIG. 9 is a graph showing the comparison of ROC curves in each sequence under noisy conditions.

Detailed Description

In order that those skilled in the art will better understand the technical solutions of the present invention and can practice the same, the present invention will be described in detail with reference to the accompanying drawings and specific examples. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby.

Example 1

The invention provides an infrared dim target detection algorithm adopting local characteristic contrast, which specifically comprises the following steps as shown in figure 2:

step 1, acquiring an original image, designing a novel nested window in order to investigate the characteristics of a neighborhood background of the original image, expanding the existing three-layer window, and effectively investigating the characteristics of the neighborhood background while retaining the advantage of single-scale calculation.

Fig. 3 shows a novel nested window designed by the present invention, which is composed of 9 × 9 sub-blocks, each of which is 3 × 3 pixel size. The middle 5 × 5 sub-blocks form an inner window (see right side of fig. 3), which is the same as fig. 1 and consists of a central sub-block T and peripheral sub-blocks SB₁～SB₁₆And intermediate subblock IB₁～IB₈And (4) forming.

Outside the nested windows, respectively with L₁₀，L₂₀，…，L₈₀Centered, each 3 x 3 subblock constitutes an outer window, identified in the same color, e.g. L₁₀，L₁₁，L₁₂，…，L₁₈An outer window L is formed₁(see fig. 3 left). Note that there is some overlap between the inner window and the outer window, e.g. SB of the inner window₁And also the outer window L₁L of₁₈。

In this new type of nested window, the algorithm will retain the advantage of single-scale computation due to the presence of the inner window isolation layer. The outer layer window has the function of observing the characteristics of the background near the target, and plays a role in further inhibiting the complex background.

Step 2, in order to accurately describe the local information of the image, aiming at the inner layer and the outer layer of the novel nested window, selecting local contrast as the characteristics of the inner layer window and the outer layer window so as to accurately describe the local information of the image, and then calculating the contrast (characteristic contrast) in a ratio difference combination mode between the two; and when the contrast is calculated, selecting the surrounding reference according to the closest filtering principle to participate in the calculation, so as to avoid the target being submerged by the surrounding highlight background.

The nested window shown in fig. 3 is placed on the original image and continuously slides, and the contrast information of each pixel point is respectively calculated at the position of each pixel point. During calculation, the algorithm adopts the idea of characteristic contrast, firstly calculates the characteristics of the inner layer window and the outer layer window respectively, and then calculates the contrast between the two. The features selected by the existing algorithm, such as variance, gradient and the like, are relatively simple and are difficult to accurately describe the local information of the image in the presence of a complex background. The step 2 specifically comprises the following steps:

step 2.1, extraction of inner layer characteristics

The characteristic of the inner window, i.e. the contrast of the inner window, should be precisely the central sub-block T and the surrounding sub-blocks SB₁～SB₁₆The contrast between them. In order to effectively deal with complex background and noise, the method firstly carries out Gaussian filtering on the central subblock and carries out nearest filtering on the surrounding subblocks before calculating the contrast ratio, and then calculates the contrast ratio by adopting a ratio-difference combination mode.

Step 2.1.1, Gaussian filtering is carried out on the central subblock

According to the matched filter theory, when a signal is processed by using a filtering method, if a filter template is consistent with the shape of the signal, the signal-to-noise ratio of the original signal can be well improved. Therefore, the invention firstly adopts a 3 multiplied by 3 two-dimensional Gaussian filter template (shown in figure 4) similar to the shape of the infrared dim target to filter the central subblock T, thereby improving the signal-to-noise ratio of the original image.

At pixel point (i, j), the matched filtering result is defined as

Where T is the center sub-block, G is the filter template in FIG. 4, IG_inTo match the filtered results.

Step 2.1.2, nearest filtering is carried out on surrounding sub-blocks

The inner window includes SB₁～SB₁₆There are 16 surrounding sub-blocks. In most contrast algorithms (including DLCM), the largest one is usually selected as the background value to participate in the contrast calculation. However, when an object appears near a highlighted background, the background selected in this manner can easily overwhelm the object. In order to alleviate the problem that the target is submerged by a highlight background, the invention selects the peripheral sub-blocks closest to the Gaussian filter value of the central sub-block to participate in calculation:

in the formula, (i, j) is the coordinate of the current pixel point, IG_inGaussian filtered value, m, representing the center subblock_SBnRepresents the mean of the nth (total of 16) surrounding sub-blocks, B_inAnd representing the final background value of the selected inner layer window.

Step 2.1.3, jointly calculating the contrast of the inner layer window by the ratio difference

The basic solving modes of the contrast comprise a ratio type, a difference type, a ratio-difference combination type and the like, and the invention adopts a ratio-difference combination calculation mode to enhance a real target and inhibit a complex background:

in the formula, RD_inThe specific difference of the inner layer window is represented as the joint contrast, xi is a smaller constraint factor and is mainly used for preventing the condition that the denominator is 0, and xi is 5 for an 8bit infrared image.

Step 2.2, extraction of outer layer characteristics

The nested windows used in the invention contain L together₁～L₈And 8 outer windows. Similar to the inner window, for the nth (total of 8) outer windows, its central sub-block is first gaussian filtered

In the formula L_n0Is the central sub-block of the nth outer window, G is the filter template in FIG. 4, IG_outnTo match the filtered results.

Then, selecting the peripheral sub-block closest to the Gaussian filter value of the central sub-block to participate in calculation

In the formula, L_niIs the ith surrounding sub-block (total of 8) of the nth outer window, m_LniMeans, IG, of the i-th surrounding sub-block of the n-th outer window_outnRepresenting the Gaussian filtered value of the central sub-block of the nth outer window, B_outnRepresenting the final background value of the nth outer window.

Thus, the ratio-difference of the nth outer window is combined with a contrast ratio of

Step 2.3, calculation of feature contrast

After the features (contrast) of the inner window and the outer window are calculated respectively, the feature contrast can be calculated between the two windows. In the algorithm of the invention, firstly, the value which is closest to the characteristics of the inner layer window in 8 outer layer windows is selected to participate in the operation according to the closest principle, namely

Then, the characteristic contrast ("contrast of contrast") between the inner window and the outer window is calculated using a joint manner of the ratio differences:

the squaring in the formula is to prevent negative values.

And 3, designing a weighting function by utilizing the isolation layer of the new nested window according to the data characteristics of the real target, weighting the characteristic contrast, further inhibiting a complex background and highlighting the target, and improving the detection efficiency.

It can be seen that in the calculation of the feature contrast, the middle sub-block IB of the inner window₁～IB₈Is not used, this partInformation is wasted. Considering that in practical application, the target generally shows the characteristic of gradually attenuating from the center to the periphery in the image, a simple weighted check function can be designed by using the information of the middle sub-block, so that the target is further highlighted, and background clutter is suppressed: and if the maximum pixel value of the center sub-block of the inner layer window is larger than the average pixel value of any middle sub-block, the weight value W (i, j) at the position is made to be 1, and otherwise, the weight value W (i, j) is made to be 0.

The final Saliency Map (salience Map, SM) will be the product of the feature contrast and the weighting function:

SM(i,j)＝RD(i,j)gW(i,j) (14)

step 4, detecting performance analysis and threshold operation

Firstly, the detection performance of the algorithm is analyzed

When the pixel point (i, j) is a real target center, a pure background (including a highlight background) region, a random noise point, a background edge, or the like, a calculation result of the algorithm of the present invention is used, as shown in fig. 5.

1. When (i, j) is the real target center, see fig. 5 (a). Assuming that the maximum gray scale of the target is P and the gray scale of the surrounding background area is Q, since the target is usually more prominent locally, P > Q is generally present, and in this case:

gauss filtered value IG of central sub-block T of inner window_inWill approach P, each surrounding sub-block SB_nAverage value m of_SBnAll close to Q, and the background value B is selected from 16 surrounding sub-blocks_inAlso close to Q, and the characteristic value RD of the inner window_inWill be relatively large and typically greater than 0.

For the outer window, in the nth (total of 8) outer windows, the center sub-block L is due to the fact that the object generally appears in a flat background_n0Gaussian filtered value IG_outnWill be close to Q, its i-th surrounding sub-block L_niAverage value m of_LniAlso close to Q, the final background value B is selected among 8 surrounding sub-blocks_outnOr close to the original value Q, so that the characteristic RD of the nth outer window_outnClose to 0, and finally selecting the final characteristic value RD in 8 outer layer windows_outAnd will also be close to 0.

Then, the weighting function is considered, and only the inner window needs to be studied, since the target generally shows a form gradually decaying from the center to the periphery, the gray value in the center sub-block T is usually the maximum, and the weighting value at this time will be 1.

Finally, after calculating and weighting the feature contrast between the inner and outer windows, it is clear that the obtained SM (i, j) will be relatively large, and generally larger than 0.

2. When (i, j) is a pure background, see fig. 5 (b). Assuming that the gray-level value of its surrounding background area is Q, then:

gauss filtered value IG of central sub-block T of inner window_inWill approach Q, each surrounding sub-block SB_nAverage value m of_SBnAlso close to Q, background value B is selected among 16 surrounding sub-blocks_inStill close to Q, so the characteristic value RD of the inner window_inWill be close to 0.

For the outer window, the final RD is similar to the outer window when (i, j) is the real target center_outAnd will also be close to 0.

For the weighting function, since the central sub-block of the inner window and each intermediate sub-block have the same gray level, the weighting value at this time will be 0.

Finally, after the feature contrast is calculated and weighted between the inner and outer windows, the obtained SM (i, j) will be close to 0, and the result is independent of the specific value of Q.

3. When (i, j) is random point noise, as shown in fig. 5(c), assuming that the noise brightness is close to the maximum brightness of the target, both are P, and the gray value of the surrounding background is Q, and P > Q, the analysis process at this time is similar to that of a real target, and SM (i, j) >0 can be obtained. However, because the point noise is usually caused by random factors such as device electrical noise, and is mostly expressed as a single pixel point, the gaussian mean operation is used in the algorithm of the present invention, so that the calculation result of the random point noise can be easily deduced, and the calculation result of the random point noise is smaller than the result of a real target, and the detection cannot be interfered.

4. When (i, j) is near the background edge. Let the gray value at one side of the edge be Q, the other side be R, and R > Q. On the bright side of the edge, see fig. 5(d), there are:

gauss filtered value IG of central sub-block T of inner window_inWill be close to R, some surrounding sub-blocks will have an average value close to Q, but other surrounding sub-blocks on the same side as the central sub-block will have an average value close to R, with a background value B selected from the 16 surrounding sub-blocks_inStill close to R, and the characteristic value RD of the inner window_inWill be close to 0.

For outer window, when some outer window is completely located at one side of edge, its characteristic RD_outnWill approach 0 and be closest to the characteristic value of the inner window, so the final characteristic value RD is finally selected from 8 outer windows_outAnd will also be close to 0.

For the weighting function, since some of the middle sub-blocks in the inner window have the same gray level as the central sub-block, the weighting value will be 0.

Finally, after calculating and weighting the feature contrast between the inner and outer windows, the resulting SM (i, j) will be close to 0.

On the dark side of the edge (not shown in the figure), a similar conclusion can be reached after a simple analysis, SM (i, j) will still be close to 0.

In addition to the above cases, a more specific case is when the object is close to the edge of a highlight background (much larger than the brightness of the real object), see fig. 5 (e). Suppose that the maximum gray of the target is P, the gray of the surrounding background area is Q, the gray of the background on the other side of the edge is R, and R > > P > Q. This time is:

gauss filtered value IG of central sub-block T of inner window_inWill be close to P, the average of some surrounding sub-blocks will be close to Q, the average of other surrounding sub-blocks will be close to R, and the background value B selected from 16 surrounding sub-blocks according to the closest principle_inWill be close to Q and the characteristic value RD of the inner window will be_inWill be relatively large and typically greater than 0.

For outer windows, some are located in the pure background region and others are located near the edge, which, based on previous analysis, isAll of them will be characterized by a near 0, final RD_outAnd will also be close to 0.

For the weighting function, since the target is slightly brighter than the surrounding neighborhood background, the weighting value at this time will be 1.

In summary, similar to the case when the real target is located in the normal background region, the obtained SM (i, j) will be larger, and is generally larger than 0.

From the above analysis, it can be seen that after the algorithm of the present invention is processed, in SM, the real target will become more prominent, and the complex background can be well suppressed. At this time, the target can be extracted by using a simple threshold operation. In the invention, SM is firstly normalized to the range of 0-1, and then the threshold is defined as:

TH＝μ+k·std (15)

in the formula, mu is the mean value of SM; std is the standard deviation of SM; k is a given coefficient, and experiments show that 30-100 is a more appropriate value range for single-target detection.

And (3) carrying out binarization operation on SM by using TH, marking the pixel points larger than TH as 1, marking the pixel points smaller than TH as 0, and finally outputting each connected region with the value of 1 as a detected target region (in order to reduce the interference of noise, firstly carrying out expansion operation on the marking result before outputting the target).

The detection results of the algorithm disclosed in this embodiment are described and verified by experiments

First, the actual detection effect of the algorithm of this embodiment in the real sequence and single frame image library

In order to verify the effectiveness of the algorithm, the algorithm is applied to seven groups of infrared image sequences under different backgrounds and a group of simulation sequences for testing (the simulation sequences comprise a small target positioned near a highlight background and are specially used for verifying the algorithm detection performance of the target near the highlight background). In addition, in order to test the performance of the algorithm in more practical scenes, the invention also uses a group of single-frame image libraries for algorithm performance test, and the image libraries comprise 23 images under different backgrounds, and only six images are taken as an example due to limited space. Table 1 and table 2 give detailed information of eight sets of image sequences and six frames of single-frame images, respectively.

TABLE 1 details of the eight groups of IR sequences

TABLE 2 detailed information of six single-frame images

Fig. 6 shows the detection of sequences using the algorithm of the present invention, each sequence being given a frame as an illustration. Fig. 7 shows the process of detecting a single frame database using the algorithm of the present invention, where six frame images are given as an illustration. As can be seen in fig. 6 and 7:

in an original image, a target is generally weak in brightness and small in size, and is not easy to find, and a lot of complex backgrounds often exist in a field, so that the target is easy to submerge; in the obtained inner layer window characteristics, the target becomes more prominent, the complex background is restrained to a certain extent, and a small amount of residue still exists; in the obtained outer layer window characteristics, the result at the uniform background is relatively smooth, and the position of the complex background has relatively large fluctuation; after the characteristic contrast is calculated between the inner layer window and the outer layer window, the target becomes more prominent, and the complex background is better inhibited; after weighting operation, the complex background is further suppressed, and the target becomes very prominent; finally, after the threshold operation, the targets in all the examples are successfully extracted, and only one false alarm appears at the dead point in the sequence 2. The effectiveness of the algorithm of the invention is proved. It is particularly noted that for sequence 8, when the target is located near a highlighted background, a weak target can still be effectively detected using the algorithm of the present invention.

Second, the detection performance of the algorithm of the present embodiment is compared with that of the existing algorithm

To further verify the advantages of the algorithm of the present invention, the present invention selects 9 existing contrast type algorithms as the comparison algorithm, including DoG, MPCM, ILCM, NLCM, RLCM, MDTDLMS, WLDM, VAR-DIFF, and DLCM algorithms. Wherein DoG and MPCM are gray scale difference type contrast algorithms, ILCM and NLCM are gray scale ratio type contrast algorithms, RLCM and MDTDLMS are gray scale difference combination type contrast algorithms, WLDM is a weighted local contrast algorithm, VAR-DIFF is a feature contrast algorithm, and DLCM is a single scale contrast algorithm.

TABLE 3 parameter values used for the comparison algorithms

Table 3 gives the parameter values used for each comparison algorithm, most of which are recommended in the original literature.

(1) First, the SCRG and BSF indexes are compared

In the invention, firstly, Signal-to-noise Ratio Gain (SCRG) and Background Suppression Factor (BSF) are selected as evaluation criteria of algorithm performance, and the evaluation criteria are defined as:

in the formula, SCR_inAnd SCR_outRespectively representing the signal-to-noise ratio of the original image and the calculated image; sigma_inAnd σ_outRespectively representing the standard deviation of the original image and the calculated image.

By comparing the SCRG value and the BSF value among the algorithms, the enhancement capability of the algorithms on the real target and the inhibition capability on the complex background can be effectively judged. Obviously, the larger the value of the two is, the better the corresponding algorithm performance is.

TABLE 4 SCRG values for different algorithms for each image sequence

TABLE 5 BSF values for different algorithms for each image sequence

TABLE 6 SCRG values for different algorithms under each single frame image

TABLE 7 BSF values for different algorithms under each single frame image

Tables 4 and 5 show the SCRG and BSF of each algorithm in 8 sequence representative frames, and tables 6 and 7 show the SCRG and BSF of each algorithm in 6 representative frames of the single frame database. From these tables, one can obtain:

the DoG is a differential algorithm, the enhancement capability of the target is not strong, the SCRG is generally low, and the BSF value is not high; MPCM is similar to the DoG case, with even worse BSF; ILCM is a ratio type algorithm, the SCRG of which is improved to a certain extent in most cases compared with the DoG and MPCM, and the BSF of ILCM is relatively good because the DoG is used as a preprocessing algorithm to eliminate partial background in advance; the SCRG of NLCM is similar to ILCM, but the BSF value is very low; the RLCM is a specific difference combined algorithm, which is better than the performance of ILCM and NLCM theoretically, but has no significant advantages in the values of SCRG and BSF due to the fact that non-negative constraint operation is not performed in the specific implementation process and more clutter residues exist in the calculation result; the MDTDLMS is more comprehensive in design, the clutter residues are suppressed by using non-negative constraint operation, and the SCRG and the BSF of the MDTDLMS are obviously improved compared with the prior algorithms; WLDM uses entropy as the weighting of contrast, and the performance is usually better than that of simple contrast algorithms such as DoG, MPCM, ILCM, NLCM and the like, but as the contrast of WLDM only adopts a difference type calculation method, the SCRG and BSF are sometimes even better than that of MDTDLMS without weighting; the VAR-DIFF uses the variance of the central area and the surrounding area as the feature, and then the contrast is obtained for the feature, the SCRG of the VAR-DIFF is obviously improved, but the BSF value of the VAR-DIFF is poor; DLCM, a new algorithm just proposed in recent years, is better in performance than many existing algorithms, including SCRG and BSF, but for sequence 8, when the target is located near the highlighted background, the SCRG of DLCM is 0, indicating that the target is submerged; the algorithm provided by the invention can obtain good SCRG and BSF under most conditions, and the capability of the algorithm in the aspects of target enhancement and background suppression is better than that of the existing algorithm.

(2) Performing ROC curve comparison

Next, for 8 infrared sequences, the invention introduces Receiver Operating Characteristic (ROC) curves to further demonstrate the effectiveness of the algorithm of the invention, as shown in fig. 8. The detection Rate (TPR) and the False alarm Rate (FPR) are defined as follows:

when the false alarm rate is fixed, the higher the detection rate is, the better the detection performance of the algorithm is; similarly, when the detection rate is fixed, the lower the false alarm rate is, the better the detection performance of the algorithm is. In the ROC curve, the farther left the curve is, the better the detection performance of the algorithm is. As can be seen in fig. 8:

the DoG and MPCM belong to the gray difference algorithm, and the performance of the algorithm is not very good in some sequences, for example, the DoG has poor performance in almost all sequences, and the MPCM has poor performance in sequences 3, 4, 5, 6, 7 and 8; the detection performances of ILCM and NLCM are relatively close to each other in a plurality of sequences by adopting a gray ratio algorithm, but the performances of ILCM and NLCM are not satisfactory for sequences 1, 2, 4, 8 and the like; RLCM and MDTDLMS are both a differential joint algorithm, the performance of the RLCM and the MDTDLMS is usually slightly improved compared with that of the previous algorithms, but the performance of the RLCM and the MDTDLMS in the sequence 4 is poor; WLDM belongs to a weighted contrast algorithm, but its performance is not very good in some sequences, e.g. sequences 1, 2, 4, 5, 6, 8, etc., since the contrast definition it employs is more primitive; VAR-DIFF belongs to a feature contrast algorithm, the performance of which is better in some sequences, but because the features used by it are simpler, the detection performance is not very good when the background is more complex, such as sequences 2, 4, 8, etc.; DLCM, as a newly proposed algorithm, can achieve good detection performance in most sequences, but for sequence 8, when the target is close to a highlight background, the performance is very poor; the algorithm provided by the invention has stable performance in 8 sequences, can obtain better performance, and is the best performance in all comparison algorithms in overall view.

Third, the algorithm of the present embodiment analyzes the computational complexity and compares it with the existing algorithm

For simplicity, assume that the resolution of the original infrared image is X Y, and the scale of each algorithm filter window or image sub-block is (2L +1)². For multi-scale algorithms, e.g. MPCM, RLCM, WLDM, etc., S represents the number of scales they use, L_iValue L, representing the ith scale_SThe value of L representing the largest dimension.

For DoG, each pixel will consume (2L +1)²Multiply by (2L +1)²A sub-addition, whereupon it is calculatedThe degree of impurity will be O (L)²XY)。

For MPCM, at each pixel location, at each scale, the averaging operation would consume (2L)_i+1)²Sub-addition and sub-division, finding 8 directional difference contrasts would consume 8 subtractions, and thus the total computational complexity at S scales would be O (SL)_S ²XY)。

ILCM and NLCM are both sub-block level algorithms, which first introduce a DoG as a pre-process, while the subsequent sub-block level operation consumes much less computation than the DoG, and thus their computation complexity is the same as that of the DoG, which is also O (L)²XY)。

RLCM requires first ordering the pixels in each sub-block, at each scale ordering would consume (2L)_i+1)²log(2L_i+1)²And (5) secondary calculation. The ratio-difference joint contrast for each subsequent direction will consume 1 division, 1 multiplication and 1 subtraction, and a total of 24 calculations for 8 directions. Thus, the final computational complexity at S scales would be O [ SL ]_S ²log(L_S ²)XY]The algorithm complexity of MDTDLMS is O (L)²XY)。

In WLDM, at each scale, the average operation per pixel will be consumed (2L)_i+1)²And (4) secondary addition. In addition, the entropy calculation requires ordering of a sub-block, which will be consumed at each scale (2L)_i+1)²log(2L_i+1)²And (5) secondary calculation. Thus, its final computational complexity at S scales will be O SL_S ²log(L_S ²)XY]。

In VAR-DIFF, the average needs to be calculated for the inner layer sub-block (2L +1)²Second addition, calculate its variance requirement (2L +1)²Subtraction, (2L +1)²Multiply by (2L +1)²And (4) secondary addition. The outer window is similar, so its computational complexity will be O (L)²XY)。

In DLCM, calculating the mean of the center subblocks will consume (2L +1)²Second addition, calculating the average of 16 surrounding sub-blocks would consume 16(2L +1)²Sub-addition, central sub-block and peripheryThe contrast of the sub-block will consume 16 subtractions and 16 comparisons. The calculation of the weighting function requires the use of the product of the mean and the difference of each of the 8 intermediate subblocks, respectively consuming 8(2L +1)²And 8 times of addition, 8 times of subtraction and 4 times of multiplication. Its computational complexity would then be O (L)²XY)。

The calculation of the algorithm of the present invention can be divided into four modules: the characteristic calculation of the inner window, the characteristic calculation of the outer window, the calculation of the characteristic contrast between the inner window and the outer window, and the application of a weighting function.

First, in the inner window feature calculation, Gaussian filtering of the center sub-block will consume (2L +1)²Second addition, the average of 16 surrounding sub-blocks will consume 16(2L +1)²The selection of the best reference for the secondary addition will consume 16 comparisons, the ratio difference between them, combined with the contrast, consumes 1 division, 1 multiplication and 1 subtraction, for a total of 17(2L +1)²+19 calculations.

Then, in the feature calculation of an outer window, Gaussian filtering of the center sub-block consumes (2L +1)²Second addition, the average of 8 surrounding sub-blocks will consume 8(2L +1)²The selection of the best reference for the second addition will consume 8 comparisons, the ratio difference between the two consumes 1 division, 1 multiplication and 1 subtraction of the contrast, 9(2L +1)²+11 calculations. Thus 8 outer windows together require 72(2L +1)²+88 calculations.

Next, in the calculation of the ratio difference joint contrast between the inner and outer windows, the best result is selected from 8 outer windows, 8 comparisons are needed, and then the ratio difference joint contrast is calculated by 1 division, 1 multiplication and 1 subtraction. The above total 11 calculations.

Finally, in the application of the weighting function, the respective average values of 8 intermediate sub-blocks in the inner window need to be calculated, and 8(2L +1) is needed in total²The second addition, followed by a comparison of the central sub-block with the 8 intermediate sub-blocks, requires 8 comparisons. Above all 8(2L +1)²+1 calculation.

In summary, for a pixel point, the algorithm of the invention needs 17(2L +1)²+19+72(2L+1)² +88+11+8(2L+1)²+1＝97(2L+1)²+119 calculations. The computational complexity of the algorithm of the present invention will then be O (L)²XY)。

TABLE 8 comparison of computational complexity for different algorithms

TABLE 9 average elapsed time comparison of different algorithms across sequences

Table 8 gives the computational complexity contrast for each algorithm. It should be noted that, the algorithm of the present invention is similar to DLCM, and a single-scale calculation window is adopted, from this viewpoint, the actual calculation amount of the algorithm of the present invention should be generally smaller than that of multi-scale algorithms such as MPCM, RLCM, and the like.

Table 9 gives the average elapsed time required to process one frame of image using different algorithms for 8 sequences. As can be seen from the table, DoG, ILCM, NLCM, etc. are several algorithms that take less time on average, but in combination with the contents of 5.2.1 and 5.2.2, it can be seen that their performance is not very good. The algorithm of the invention can effectively improve the detection performance and simultaneously has acceptable time consumption.

Note that, here, a PC platform, serial processing architecture is used. In fact, the algorithm has better parallel processing potential, and in practical engineering application, the algorithm can be realized by using some parallel processing platforms so as to further improve the real-time performance of the algorithm. For example, in the algorithm of the present invention, the inner window and the outer window can be processed in parallel, the central sub-block and the surrounding sub-blocks can be processed in parallel, and so on.

Fourth, analysis of noise immunity of the algorithm of the present embodiment

Noise is an important factor that interferes with image quality, reducing the target detection rate. To test the resistance of the algorithm of the present invention to noise, 8 sequences were used, to which 0-mean white gaussian noise of different variance was added, respectively, and then the performance of the algorithm under noisy conditions was tested, shown in ROC, see fig. 9. It can be seen from the figure that the detection performance of the algorithm is basically stable or only slightly reduced after noise is added in the sequence, which proves that the algorithm of the invention has better anti-noise capability.

Aiming at the problem of infrared small and weak target detection, the invention firstly provides a novel nested window, on one hand, the advantage that small targets with different scales can be detected through single-scale calculation is reserved, and on the other hand, the characteristics of the neighborhood background can be fully inspected.

Secondly, in order to accurately describe the local information of the image, the local contrast is selected as the characteristics of the inner layer window and the outer layer window, the characteristic contrast is calculated between the inner layer window and the outer layer window, and the reference in the contrast calculation is selected by adopting the closest principle so as to avoid that the target is submerged when being close to the highlight background.

And finally, according to the data characteristics of the real target, an isolation layer of the nested window is utilized to provide a weighting function which is simple in calculation and very effective so as to improve the detection efficiency.

Experiments in a plurality of real and simulation sequences and images show that compared with the existing algorithm, the algorithm provided by the invention can obtain better target enhancement and background suppression effects, has obvious advantages in the aspects of detection rate and false alarm rate, and only needs sub-second time for averagely processing one frame of image.

The above-mentioned embodiments are only preferred embodiments of the present invention, and the scope of the present invention is not limited thereto, and any simple modifications or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the technical scope of the present invention are included in the scope of the present invention.

Claims

1. An infrared weak and small target detection algorithm adopting local characteristic contrast is characterized by comprising the following steps:

2. The algorithm for detecting infrared weak and small targets with local feature contrast as claimed in claim 1, wherein the new nested window is composed of 9 x 9 sub-blocks in total, and each sub-block has a size of 3 x 3 pixel points; wherein, the middle 5 × 5 sub-blocks form an inner layer window, and comprise a central sub-block T and peripheral sub-blocks SB₁～SB₁₆And intermediate subblock IB₁～IB₈Composition is carried out;

3. The algorithm for detecting the infrared weak and small target by using the local feature contrast as claimed in claim 2, wherein the step 2 comprises the following steps:

step 2.1, extraction of inner layer characteristics

Step 2.1.1, Gaussian filtering is carried out on the central subblock

at the pixel point (i, j), the matched filtering result is defined as:

step 2.1.2, nearest filtering is carried out on surrounding sub-blocks

step 2.2, extraction of outer layer characteristics

step 2.3, calculation of feature contrast

4. the algorithm for detecting infrared weak and small objects by using local feature contrast as claimed in claim 3, wherein in the step 3, only when the maximum pixel value of the center sub-block of the inner window is greater than the average pixel value of any intermediate sub-block, the weight W (i, j) at the position is made to be 1, otherwise, the weight W (i, j) is made to be 0;

SM(i,j)＝RD(i,j)gW(i,j) (14)

5. The algorithm for detecting the infrared weak and small target by using the local feature contrast as claimed in claim 4, wherein the step 4 comprises the following steps:

TH＝μ+k·std (15)