CN109816641B - Multi-scale morphological fusion-based weighted local entropy infrared small target detection method - Google Patents

Abstract

The invention provides a weighted local entropy infrared small target detection method based on multi-scale morphological image fusion, which comprises the steps of firstly, converting an infrared image into a gray domain for processing; secondly, carrying out multi-scale morphological Top-Hat image segmentation processing on the infrared image, solving the image difference on the basis of Top-Hat of adjacent scales to obtain a minimum difference image, and fusing the minimum difference image and the minimum mean value image of the image subjected to Top-Hat transformation to obtain an image subjected to background suppression; then, a local entropy information graph is obtained by calculating the local entropy of the initial image; then, performing dot multiplication on the image subjected to background suppression and the local entropy information image, and normalizing to obtain a saliency map of the infrared small target; and finally, filtering and binarizing the infrared small target saliency map by using a threshold segmentation technology to obtain a processed image, wherein the area with the binarization value of 1 is the infrared small target. The method is suitable for the field of infrared small target detection, can effectively improve the accuracy of infrared small target detection, and effectively reduces the false alarm rate.

Description

Multi-scale morphological fusion-based weighted local entropy infrared small target detection method

Technical Field

The invention belongs to the technical field of digital image processing, and particularly relates to a weighted local entropy infrared small target detection method based on multi-scale morphological fusion.

Background

The infrared small target detection technology has become an important branch in the fields of national defense, military and civil use, and has the advantages of all weather, long detection distance, good hiding performance and the like, so that the infrared small target detection technology is widely applied to the fields of infrared long-distance target tracking, infrared terminal guidance, border national defense, medical treatment and health care and the like. However, because the imaging sensor of the infrared image is influenced by factors such as the temperature and the material of the device and the inherent characteristics of the infrared small target image, in a single-frame image, the small target pixel in the infrared small target imaging has small occupation ratio, weak signal and easy interference, and finally, the target detection and tracking become extremely difficult. Therefore, infrared small target detection becomes a hotspot and difficulty of research in the field of digital image processing.

With the rapid development of the information technology field, the scale of image data becomes larger and larger, and how to complete image analysis quickly and accurately in the face of such huge and numerous data has become an important branch of current research. Since 1998 Laurent Itti proposed a visual saliency detection method, visual saliency detection technology has gained increasing attention from scholars. Visual saliency detection is a problem that arises in the field of Computer-Vision (Computer-Vision) to mimic the visual attention mechanism of primates. In recent years, scholars propose a plurality of related algorithms based on the mechanism and apply the algorithms to the field of infrared small target detection and tracking. The prominent characteristic of the infrared small target is extracted based on a visual attention mechanism and used for detecting the infrared small target, so that a good detection result can be obtained. Chen et al propose a Local Contrast Measurement (LCM) algorithm based on Human Visual Contrast mechanism (HVS). The LCM algorithm highlights the saliency of the small targets, inhibits background clutter and obtains a saliency map of the infrared small target image according to the difference of Local Contrast (LC) characteristics of the targets and the background. Han et al propose an improved LCM algorithm, called Improved Local Contrast Method (ILCM), which is also based on the HVS algorithm. The LCM algorithm is improved, the calculation speed is high, the false alarm rate is reduced, and the detection under the complex background is still a difficult problem.

Disclosure of Invention

Aiming at the defects of the prior art in the field of infrared Image small target detection under a complex background, the invention provides an infrared small target detection algorithm based on Weighted Local Entropy (WLEMMF) of Multi-scale Morphological Image Fusion.

The algorithm idea for realizing the invention is as follows: firstly, converting an infrared image into a gray domain for processing; secondly, according to the proposed WLEMMF algorithm, carrying out multi-scale morphological Top-Hat image segmentation processing on the infrared image, solving the image difference on the basis of Top-Hat of adjacent scales to obtain a minimum difference image, and fusing the minimum difference image and the minimum mean value image of the image subjected to Top-Hat transformation to obtain an image subjected to background suppression; then, a local entropy information graph is obtained by calculating the local entropy of the initial image; then, performing dot multiplication on the image subjected to background suppression and the local entropy information image, and normalizing to obtain a saliency map of the infrared small target; and finally, filtering and binarizing the infrared small target saliency map by using a threshold segmentation technology to obtain a processed image, wherein a bright spot area (an area binarized into 1) is the infrared small target.

Based on the principle, the technical scheme of the invention is as follows:

the method for detecting the weighted local entropy infrared small target based on the multi-scale morphological fusion is characterized by comprising the following steps of: the method comprises the following steps:

step 1: inputting an infrared image Img to be detected, wherein the size of the image is m x n;

step 2: converting the input image Img into a gray level image to obtain an image Img_in；

And step 3: the following steps were used for background suppression:

step 3.1: setting I morphological pretreatment structural units m_iI is more than or equal to 1 and less than or equal to I, wherein I is an integer greater than zero, and the size of the morphological pretreatment structural unit increases with the increase of I;

step 3.2: selecting morphological pretreatment building Block m_iFor image Img_inPerforming Top-Hat morphological image preprocessingAfter the preprocessing, the image is recorded as WTHS_i；

Step 3.3: repeating the step 3.2 until all the set morphological pretreatment structural units and the image Img_inFinishing Top-Hat morphological image preprocessing to obtain a multiscale Top-Hat preprocessing image set WTHS;

step 3.4: and (3) carrying out difference on gray values of corresponding pixel points of two adjacent images in the multi-scale Top-Hat preprocessed image set WTHS to obtain a difference graph:

WTHS_{Top_Hat(i+1)}＝WTHS_i+1-WTHS_i，1≤i≤I-1

step 3.5: repeating the step 3.4 to obtain a difference image set WTHS formed by difference images between two adjacent images of the group I-1 in the multi-scale Top-Hat preprocessing image set WTHS_{Top_Hat}；

Step 3.6: for difference atlas WTHS_{Top_Hat}Performing minimum fusion of element gray values to obtain a fusion image WTHS_{Top_Hat_min}(ii) a Calculating the gray level mean value of each image in the multiscale Top-Hat preprocessing image set WTHS, and taking the minimum mean value image WTHS_mean(min)；

Step 3.7: image WTHS_{Top_Hat_min}And minimum mean value diagram WTHS_mean(min)Adding gray values of corresponding pixels in the image to obtain a fusion image Img with suppressed background_WTH；

And 4, step 4: calculating an image Img_inInformation entropy of local gray of image of (1):

step 4.1: selecting an information entropy calculation window pair image Img with the radius of M_inPerforming local gray information entropy operation from top to bottom and from left to right: center slave image Img of information entropy calculation window_inStarts sliding, and calculates the local information entropy under the window until sliding to the image Img_inUp to the (M-M, n-M) position of (A), image Img_inThe entropy filling of the edge information is 0, and the diagram of the entropy of the local gray level information is obtained and recorded as Img_entropy；

Step 4.2: img local gray level information entropy diagram_entropyAnd fused graph Img_WTHMultiplying the pixel values of the corresponding pixels and classifyingObtaining a weighted local gray level information entropy chart Img after normalization_{L_entropy}As a saliency map of infrared small targets;

and 5: calculating a gray information entropy self-adaptive threshold:

step 5.1: calculating the infrared small target saliency map Img_{L_entropy}And find the maximum value H of the pixel values therein_max；

Step 5.2: calculating threshold value T ═ α (H)_max-Mean)+βMean²+ ε Var, where α, β, ε are set normal numbers;

step 6: utilizing the threshold T obtained in the step 5 to obtain the saliency map Img of the infrared small target_{L_entropy}Carrying out binarization processing, wherein the pixel value in the image is set to be 0 when the pixel value is lower than T, and is set to be 1 when the pixel value is not lower than T; obtaining a binarization result Img of the infrared small target saliency map_{L_entropy_B}And as a final infrared small target distribution detection image, the area with the binary value of 1 is the infrared small target.

Advantageous effects

Compared with the prior art, the invention has the following advantages:

1) providing a new algorithm WLEMMF, processing images by multi-scale morphology Top-Hat, calculating image difference of the images obtained by different scales, fusing the images to obtain a minimum difference image of each pixel on the basis of the difference images under each scale, and obtaining an enhanced morphology preprocessing background suppression image as a new weight of image gray information entropy;

2) selecting a proper radius of a scale window, calculating local gray information entropy, fusing an enhanced morphological processing algorithm and a local gray information entropy graph, inhibiting a large-area background, and effectively retaining small target information to obtain a small target saliency graph;

3) and the final saliency map is formed by weighting, a self-adaptive image threshold algorithm is introduced, the information of the saliency map is highlighted in a binarization mode, and the infrared small target saliency map is finally obtained.

Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description, or may be learned by practice of the invention.

Drawings

The above and/or additional aspects and advantages of the present invention will become apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings of which:

fig. 1 is a basic block diagram of infrared weak and small target detection.

Fig. 2 shows the results of infrared weak and small target detection in example 1: in the first row, a first column is an infrared original input image; the second column is a significant image obtained after the algorithm processing; the third column is the detection result of the adaptive algorithm. In the second row, a first column is a space diagram of the infrared original input image; the second column is a spatial map of a saliency image obtained after the algorithm processing; the third column is a space diagram of the detection result of the adaptive algorithm.

FIG. 3 shows an example of the result of detection of infrared small and weak targets 2;

FIG. 4 shows an example of the result of detection of infrared small and weak targets 3;

FIG. 5 shows an example of the result of infrared weak and small target detection 4;

FIG. 6 shows an example of the results of detection of infrared small and weak targets 5;

fig. 7 shows an example 6 of the result of detection of infrared weak and small targets.

Detailed Description

The following detailed description of embodiments of the invention is intended to be illustrative, and not to be construed as limiting the invention.

Referring to fig. 1, the steps implemented by the present invention are as follows:

step 1: inputting an infrared image Img to be detected, wherein the size of the image is m x n.

Step 2: converting the input image Img into a gray level image to obtain an image Img_in。

And step 3: background suppression:

based on basic operation of expansion and corrosion in morphology, f (i, j) is set as an initial gray image, and m (x, y) is a structural unit of expansion and corrosion. Performing dilation operation on the initial gray image by using the structural unit m

Is defined as:

in the formula, D_f、D_mThe fields of definition for f (i, j) and m (x, y), respectively. Corrosion note f Θ m, defined as:

(fΘm)(i，j)＝min{f(i-x，j-y)-m(x，y)|(i-x，j-y)∈D_f，(x，y)∈D_m}

dilation and erosion are fundamental operations of morphology, dilation can be a local maximum filtering operation, and erosion can be a local minimum filtering operation. On the basis of this, an image on operation f omicron and an image off operation f · m are derived, which are defined as:

the open operation is to perform corrosion on the initial image by the structural unit m (x, y) and then perform expansion operation on the result image, so that encouragement points can be restrained to a certain extent, and the close operation is to perform corrosion operation on the result image by the structural unit m (x, y) and then perform expansion operation on the initial image.

The initial image is subtracted by the result of the open operation, which is defined as Top-Hat transform (Top-Hat), and the algorithm can obtain the points of the edges of some regions in the image, which provides a thought for finding small targets in the gray-scale image. Is defined as:

G＝f-(fοm)

Top-Hat at a single scale cannot effectively remove isolated noise in an image.

The algorithm improves the traditional morphological Top-Hat algorithm,an enhanced morphological Top-Hat is formed. The algorithm carries out Top-Hat transformation under multiple scales aiming at the initial image, and prefilter images WTHS obtained under different scales (the scales are from small to large) are processed_iWherein i is a scale index, and Top-Hat under adjacent scales are fused to obtain an image difference atlas WTHS_{Top_Hat}. Fusing the obtained difference maps to obtain the minimum difference map WTHS_{Top_Hat_min}. And finding the graph with the minimum mean value and the WTHS under the image difference graph_{Top_Hat_min}Performing image addition operation for fusion to obtain an image Img_WTH。

According to the specification of the Society of Optical engineering (SPIE), an infrared image with the size of m × n has the number of infrared small target pixel points not more than 0.12% of the pixel points in the image. Setting I morphological pretreatment structural units, wherein I is an integer larger than zero. In the computer field m_iThe indicated size is (2k +1)²Square region of (2), selecting the largest structure m in the set_IDimension of (2k +1)²≤0.0012*m*n。

Wherein I is more than or equal to 1 and less than or equal to I, and m is used for reducing the omission factor_iThe above conditions need to be met, so the scale selection can be set according to the prior knowledge of the size of the input image.

Step 3.1: setting I morphological pretreatment structural units m_iI is more than or equal to 1 and less than or equal to I, wherein I is an integer greater than zero, and the size of the morphological pretreatment structural unit increases with the increase of I;

step 3.2: selecting morphological pretreatment building Block m_iFor image Img_inPreprocessing the Top-Hat morphological image to obtain a preprocessed image, and recording the preprocessed image as WTHS_i；WTHS_i＝Img_in-(Img_inοm_i)，Img_inIs a grey scale map of the input image, 'omicron' is the on operation in morphology;

step 3.3: repeating the step 3.2 until all the set morphological pretreatment structural units and the image Img_inFinishing Top-Hat morphological image preprocessing to obtain a multiscale Top-Hat preprocessing image set WTHS;

step 3.4: and (3) carrying out difference on gray values of corresponding pixel points of two adjacent images in the multi-scale Top-Hat preprocessed image set WTHS to obtain a difference graph:

WTHS_Top-Hat(i+1)＝WTHS_i+1-WTHS_i，1≤i≤I-1

step 3.5: repeating the step 3.4 to obtain a difference image set WTHS formed by difference images between two adjacent images of the group I-1 in the multi-scale Top-Hat preprocessing image set WTHS_{Top_Hat}；

Step 3.6: for difference atlas WTHS_{Top_Hat}Performing minimum fusion of element gray values, namely extracting the minimum value of the gray value of each pixel point corresponding to each image in the difference image set, fusing the minimum values into a new image to obtain a fused image WTHS_{Top_Hat_min}(ii) a Calculating the gray average value of each image in the multi-scale Top-Hat preprocessing image set WTHS, namely the pre-filtering image WTHS under the ith scale_iMean value of WTHS_mean(i)Selecting the minimum mean value graph and recording as WTHS_mean(min)；

Step 3.7: image WTHS_{Top_Hat_min}And minimum mean value diagram WTHS_mean(min)Adding gray values of corresponding pixels in the image to obtain a fusion image Img with suppressed background_WTH；

Img_WTH(j，k)＝WTHS_{Top_Hat_min(j，k)}+WTHS_{mean(min)(j，k)}

Where the subscripts (j, k) are the coordinates of the corresponding pixels.

And 4, step 4: calculating an image Img_inInformation entropy of local gray of image of (1):

in one frame of infrared image, the small infrared target can be regarded as a singular point which exists in the background and is not compatible with the background, and the characteristic is helpful for separating the small target from the complex background and detecting the small target. However, when the background is complex and the noise interference is strong, the target information is easily submerged, which causes that the small target is difficult to separate from the background and the detection cannot determine the real target. The image gray information entropy is a statistic of information stray characteristics, reflects the average information quantity of an image, and represents the aggregation characteristics of image gray distribution. And the small target can be prevented from being missed to be detected to a greater extent by selecting a proper window function.

The gray information entropy in the image is slightly different from the extracted gray distribution information of the image according to the difference of the window function size, and the smaller the window is, the smaller the small target information in the gray information can be better extracted. However, the window is often too small, which results in excessive information extraction, and some isolated noise points are effectively counted, so that not only is the calculation complexity increased, but also the robustness of small target detection is reduced, therefore, a proper size M of the entropy window needs to be selected, and the window radius is set to be 2. And calculating the gray information entropy of the field, replacing the pixel value of the window center position with the entropy, and obtaining a local entropy information graph of the image after traversing the whole image.

Step 4.1: selecting radius as M, window size as (2M +1)²To the image Img_inPerforming r times of local gray scale information entropy operation from top to bottom and from left to right;

0≤r≤(m-2M)*(n-2M)

setting the window space to omega_θCenter of information entropy calculation window from image Img_inStarts sliding, and calculates the local information entropy under the window until sliding to the image Img_inAt (M-M, n-M), every time the window center moves by one pixel point, all statistical information needs to be recalculated; image Img_inThe entropy filling of the edge information is 0, and the diagram of the entropy of the local gray level information is obtained and recorded as Img_entropy。

Defining (x, y) as the coordinate of any pixel point passing through the center of the window in the sliding process of the entropy window, taking the coordinate as the center pixel point of the window, and setting (i, j) to be belonged to the space omega_θInner pixel point coordinates, f_(i，j)Representing the image gray value of the change point.

The gray information entropy of the corresponding center pixel:

let ε equal to 10E-6, wherein

P_(i，j)Represents the region omega_θThe gray level spatial distribution of each point:

P_(i，j)＝f_(i，j)/N²

wherein N is²Representing the size of the region, i.e. (2M +1)². The area gray level mean value is used as the entropy calculation of the image information, and the gray level distribution situation on the space can be reflected reliably.

Initially (x, y) — (M +1) is defined, and the local information entropy is calculated as Z_(x，y)Construction of size and Img_inSame image local information entropy map, fill in entropy value Z at the above (x, y) coordinates_(x，y)Obtaining partial image information entropy diagram

And then traversing all coordinates (x, y) in turn according to the movement of the center of the entropy window function, wherein x belongs to (M +1, M-M), y belongs to (M +1, n-M), and the rest edge points are filled with 0 in the information entropy diagram to obtain a complete image information entropy diagram Img_entropy。

Step 4.2: fusing the picture Img_WTHPixel value of each pixel point is used as information entropy chart Img_entropyThe local gray level information entropy diagram Img_entropyAnd fused graph Img_WTHPerforming point multiplication operation of the image, namely multiplying pixel values of corresponding pixel points, and normalizing to obtain a weighted local gray level information entropy chart Img_{L_entropy}As a saliency map of infrared small targets.

And 5: calculating a gray information entropy self-adaptive threshold:

in step 4, a saliency map of the infrared small target is calculated, which is also the most critical part of the whole algorithm. The key link in the image segmentation method of visual saliency is to calculate the saliency of each element in an image and calculate and give a saliency adaptive threshold value, so that the aim of image segmentation is fulfilled.

Step 5.1: calculating the infrared small target saliency map Img_{L_entropy}And find the maximum value H of the pixel values therein_max；

H_max＝max{(gray_(i，j)}

Wherein gray_(i，j)Significance map Img representing infrared small target_{L_entropy}The pixel value of each pixel point.

Step 5.2: after the infrared small target saliency map is obtained, binarization is performed because the contrast is low. Calculating threshold value T ═ α (H)_max-Mean)+βMean²+ ε Var, where α, β, ε are the set normal numbers,

step 6: utilizing the threshold T obtained in the step 5 to obtain the saliency map Img of the infrared small target_{L_entropy}Carrying out binarization processing, wherein the pixel value in the image is set to be 0 when the pixel value is lower than T, and is set to be 1 when the pixel value is not lower than T; obtaining a binarization result Img of the infrared small target saliency map_{L_entropy_B}And as a final infrared small target distribution detection image, the area with the binary value of 1 is the infrared small target.

Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention, and that variations, modifications, substitutions and alterations can be made in the above embodiments by those of ordinary skill in the art without departing from the principle and spirit of the present invention.

Claims (1)

1. A weighted local entropy infrared small target detection method based on multi-scale morphological fusion is characterized by comprising the following steps: the method comprises the following steps:

step 1: inputting an infrared image Img to be detected, wherein the size of the image is m x n;

step 2: converting the input image Img into a gray level image to obtain an image Img_in；

And step 3: the following steps were used for background suppression:

step 3.1: setting I morphological pretreatment structural units m_iI is more than or equal to 1 and less than or equal to I, wherein I is an integer greater than zero, and the size of the morphological pretreatment structural unit increases with the increase of I;

step 3.2: selecting morphological pretreatment building Block m_iFor image Img_inPreprocessing the Top-Hat morphological image to obtain a preprocessed image, and recording the preprocessed image as WTHS_i；

Step 3.3: repeating the step 3.2 until all the set morphological pretreatment structural units and the image Img_inFinishing Top-Hat morphological image preprocessing to obtain a multiscale Top-Hat preprocessing image set WTHS;

step 3.4: and (3) carrying out difference on gray values of corresponding pixel points of two adjacent images in the multi-scale Top-Hat preprocessed image set WTHS to obtain a difference graph:

WTHS_{Top_Hat(i+1)}＝WTHS_i+1-WTHS_i，1≤i≤I-1

step 3.5: repeating the step 3.4 to obtain a difference image set WTHS formed by difference images between two adjacent images of the group I-1 in the multi-scale Top-Hat preprocessing image set WTHS_{Top_Hat}；

Step 3.6: for difference atlas WTHS_Top-HatPerforming minimum fusion of element gray values to obtain a fusion image WTHS_Top-Hat-min(ii) a Calculating the gray level mean value of each image in the multiscale Top-Hat preprocessing image set WTHS, and taking the minimum mean value image WTHS_mean(min)；

Step 3.7: image WTHS_{Top_Hat_min}And minimum mean value diagram WTHS_mean(min)Adding gray values of corresponding pixels in the image to obtain a fusion image Img with suppressed background_WTH；

And 4, step 4: calculating an image Img_inInformation entropy of local gray of image of (1):

step 4.1: selecting an information entropy calculation window pair image Img with the radius of M_inPerforming local gray information entropy operation from top to bottom and from left to right: center slave image Img of information entropy calculation window_inStarts sliding, and calculates the local information entropy under the window until sliding to the image Img_inUp to the (M-M, n-M) position of (A), image Im_ginThe entropy filling of the edge information is 0, and the diagram of the entropy of the local gray level information is obtained and recorded as Img_entropy；

Step 4.2: img local gray level information entropy diagram_entropyAnd fused graph Img_WTHMultiplying pixel values of corresponding pixel points, and normalizing to obtain a weighted local gray level information entropy chart Img_{L_entropy}As a saliency map of infrared small targets;

and 5: calculating a gray information entropy self-adaptive threshold:

step 5.1: calculating the infrared small target saliency map Img_{L_entropy}And find the maximum value H of the pixel values therein_max；

Step 5.2: calculating threshold value T ═ α (H)_max-Mean)+βMean²+ ε Var, where α, β, ε are set normal numbers;

step 6: utilizing the threshold T obtained in the step 5 to obtain the saliency map Img of the infrared small target_{L_entropy}Carrying out binarization processing, wherein the pixel value in the image is set to be 0 when the pixel value is lower than T, and is set to be 1 when the pixel value is not lower than T; obtaining a binarization result Img of the infrared small target saliency map_{L_entropy_B}And as a final infrared small target distribution detection image, the area with the binary value of 1 is the infrared small target.

Images