Infrared small target detection method based on gray scale difference and information entropy

Abstract

The invention discloses an infrared small target detection method based on gray scale difference and information entropy, and belongs to the technical field of infrared image detection. The method solves the problems that features selected by an existing method are insufficient in capability of inhibiting complex backgrounds, and the detection accuracy of infrared small target detection based on the features selected by the existing method is low. According to the gradient domain feature extraction method, the extracted gradient mode features can make the target and other areas obviously distinguished, the target area can be enhanced by extracting the direction angle features and combining filtering, background area suppression is achieved, the local information entropy serves as a weight factor, and information loss caused in the gradient domain feature extraction process can be effectively supplemented. And finally, by integrating the filtered direction angle feature map, the processed gradient mode feature map and the entropy weight feature map, the accuracy of infrared small target detection can be improved, and the universality of multiple use environments is improved. The method can be applied to infrared small target detection.

Description

Technical Field

[0001] The invention belongs to the technical field of infrared image detection, and in particular relates to an infrared small target detection method based on grayscale difference and information entropy. Background Art

[0002] With the continuous development of infrared technology, the use scenarios of various infrared detection devices are more abundant. As a passive detection method, infrared detection has stronger concealment. The infrared detection system is less affected by light, has strong penetration ability, can implement all-weather reconnaissance and early warning detection, and can cope with the influence of bad weather such as smoke, fog and haze. Infrared technology is widely used in civil and military fields, including air and sea target monitoring, reconnaissance and early warning, and search and rescue. When the target to be measured is far away from the infrared detection system, the target usually exists in the form of a small target in the imaging picture, so it is necessary to detect the infrared small target image accurately and quickly. According to the definition of the international organization SPIE, targets whose area does not exceed 0.12% of the entire image are classified as small targets. On the one hand, small infrared targets have the characteristics of weak energy, small size, few pixels, and unclear texture features; on the other hand, due to different application scenarios, the influence of complex background and noise interference also makes the detection of small infrared targets more difficult.

[0003] Existing research on infrared small target detection mainly focuses on single-frame images and multi-frame sequence images. At present, multi-frame image detection is mainly divided into two categories. One is the association verification type, that is, first use a single-frame image to find several possible targets respectively, and then use the continuity of the target's motion trajectory to check and judge in the multi-frame image to extract the motion trajectory of the real target; the other is the direct solution method, that is, use the grayscale change information in the time domain to directly extract the target in the multi-frame image. Faced with high-speed and flexible motion devices, the target runs fast, the background changes rapidly, and the system is difficult to detect continuously. It is difficult to obtain a complete sequence-level infrared detection image. On the other hand, a high-precision single-frame image detection algorithm is the basis and premise of multi-frame image target detection. However, the infrared small target detection method of a single-frame image is based on artificial feature selection, and the detection task is mainly carried out around the image space-frequency characteristics, local contrast characteristics, and low-rank sparse characteristics. For example, the patent application with publication number CN118521768A discloses a "small infrared target detection method based on non-window structure local contrast measurement". This method uses the gradient characteristics of the image to detect the target through mean filtering and gradient operators, but it is difficult to filter out pixel-level highlight clutter and strong edges, so the ability to suppress complex backgrounds is still obviously insufficient. The patent application with publication number CN118397302A discloses a "small infrared target detection method based on differential directional hybrid sparse constraints combined with significant coherence enhancement". This method generates significant coherence consistency factors by processing multiple infrared images and then performs detection, but the model relies too much on the time and space information in the infrared image data, which can easily lead to background reduction and target loss.

[0004] In summary, the features selected by the existing methods still have the problem of insufficient ability to suppress complex backgrounds. When infrared small targets are detected based on the features selected by the existing methods, there is still the problem of low detection accuracy. Therefore, it is necessary to propose a new infrared image feature extraction method to suppress complex backgrounds and improve detection accuracy. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the features selected by the existing methods have insufficient ability to suppress complex backgrounds, and the detection accuracy of infrared small targets obtained by detecting the features selected by the existing methods is low, and a method for detecting infrared small targets based on grayscale difference and information entropy is proposed.

[0006] The technical solution adopted by the present invention to solve the above technical problems is: a method for detecting small infrared targets based on grayscale difference and information entropy, the method specifically comprising the following steps:

[0007] Step S1, preprocessing the original infrared image based on the Scharr gradient operator, and obtaining a gradient modulus feature map and an angular feature map of the original infrared image according to the preprocessing result;

[0008] Step S2, constructing a direction angle filtering template, filtering the direction angle feature map through the direction angle filtering template, and extracting the filtered direction angle feature map of the original infrared image;

[0009] Step S3, constructing a dual-neighborhood block local contrast module, wherein the dual-neighborhood block local contrast module includes a central reference area, a middle buffer area and an outer background area;

[0010] The dual-neighborhood block local contrast module is used to process the gradient mode feature map of the original infrared image to obtain a processed gradient mode feature map;

[0011] Step S4, respectively calculating the entropy weight feature of each pixel in the original infrared image to obtain an entropy weight feature map of the original infrared image;

[0012] Step S5, integrating the filtered azimuth feature map, the processed gradient modulus feature map, and the entropy weight feature map of the original infrared image to obtain a final feature map of the original infrared image;

[0013] Step S6: Detect small targets in the original infrared image based on the obtained final feature map.

[0014] Furthermore, the specific process of step S1 is as follows:

[0015] Step S11, using the Scharr gradient operator in the horizontal direction to perform convolution calculation on the original infrared image, respectively obtaining the horizontal gradient of each pixel point, and using the Scharr gradient operator in the vertical direction to perform convolution calculation on the original infrared image, respectively obtaining the vertical gradient of each pixel point;

[0016] Step S12: for any pixel point, calculate the gradient modulus feature and the direction angle feature of the pixel point according to the horizontal gradient and the vertical gradient of the pixel point;

[0017] Similarly, after traversing each pixel point in the original infrared image, the gradient modulus feature map and the directional angle feature map of the original infrared image are obtained.

[0018] Furthermore, the Scharr gradient operator in the horizontal direction and the Scharr gradient operator in the vertical direction are respectively:

[0019]

[0020] Among them, G x_newis the Scharr gradient operator in the horizontal direction;

[0021]

[0022] Among them, G y_new is the Scharr gradient operator in the vertical direction.

[0023] Furthermore, in step S12, the calculation method of the gradient modulus feature and the directional angle feature of the pixel point is specifically as follows:

[0024] Taking the pixel point (x, y) as an example, the horizontal gradient of the pixel point (x, y) is recorded as The vertical gradient of the pixel point (x, y) is recorded as Then the gradient modulus feature of the pixel point (x, y) is:

[0025]

[0026] Among them, M(x,y) is the gradient modulus feature of the pixel point (x,y);

[0027] The direction angle feature of the pixel point (x, y) is:

[0028]

[0029] Among them, D(x,y) is the direction angle feature of the pixel point (x,y).

[0030] Furthermore, the size of the directional angle filter template is 3×3, the elements of the first row and first column in the directional angle filter template are π / 4, the elements of the first row and second column are π / 2, the elements of the first row and third column are 3π / 4, the elements of the second row and first column are 0, the elements of the second row and second column are 0, the elements of the second row and third column are π, the elements of the third row and first column are -π / 4, the elements of the third row and second column are -π / 2, and the elements of the third row and third column are -3π / 4.

[0031] Furthermore, in step S2, the directional angle feature map is filtered by using the directional angle filter template to extract the filtered directional angle feature map of the original infrared image. The specific process is as follows:

[0032] Slide the directional angle filter template on the directional angle feature map so that the center of the directional angle filter template passes through each pixel point on the directional angle feature map in sequence;

[0033] For the pixel point (x, y) on the angular feature map, the angular feature map composed of (x, y) and all the pixels in the 3×3 neighborhood of (x, y) is recorded as D′, and the angular feature map D′ is convolved with the angular filter template, and the convolution result is used as the filtered angular feature of the pixel point (x, y):

[0034] GD(x,y)=D′*DFT

[0035] Where * represents convolution operation, DFT represents azimuth filter template, and GD(x,y) is the azimuth feature of pixel point (x,y) after filtering.

[0036] The filtered angular features of all pixels constitute the filtered angular feature map of the original infrared image.

[0037] Furthermore, the size of the dual-neighborhood block local contrast module is 15×15; wherein:

[0038] The area consisting of pixels located in the first to third rows and the first to third columns in the dual-neighborhood block local contrast module is denoted as B. o1 ;

[0039] The area consisting of pixels located in the first to third rows and in the fourth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B o2 ;

[0040] The area consisting of pixels located in the first to third rows and the thirteenth to fifteenth columns in the dual-neighborhood block local contrast module is denoted as B. o3 ;

[0041] The area consisting of pixels located in the fourth to twelfth rows and the thirteenth to fifteenth columns in the dual-neighborhood block local contrast module is denoted as B o4 ;

[0042] The area consisting of pixels located in the 13th to 15th rows and the 13th to 15th columns in the dual-neighborhood block local contrast module is denoted as B. o5 ;

[0043] The area consisting of pixels located in the 13th to 15th rows and the 4th to 12th columns in the dual-neighborhood block local contrast module is denoted as B o6 ;

[0044] The area consisting of pixels located in the thirteenth to fifteenth rows and in the first to third columns in the dual-neighborhood block local contrast module is denoted as B o7 ;

[0045] The area consisting of pixels located in the fourth to twelfth rows and in the first to third columns in the dual-neighborhood block local contrast module is denoted as B o8 ;

[0046] The area consisting of pixels located in the fourth to sixth rows and the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i1 ;

[0047] The area consisting of pixels located in the fourth to sixth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as B i2 ;

[0048] The area consisting of pixels located in the fourth to sixth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i3 ;

[0049] The area consisting of pixels located in the seventh to ninth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i4 ;

[0050] The area consisting of pixels located in the tenth to twelfth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i5 ;

[0051] The area consisting of pixels located in the tenth to twelfth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as B. i6 ;

[0052] The area consisting of pixels located in the tenth to twelfth rows and in the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i7 ;

[0053] The area consisting of pixels located in the seventh to ninth rows and in the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i8 ;

[0054] The area consisting of pixels located in the seventh to ninth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as T;

[0055] Region B o1 , B o2 , B o3 , B o4 , B o5 , B o6 , B o7 and B o8 Composing the outer background area, area B i1 , B i2 , B i3 , B i4 , B i5 , B i6 , B i7 and B i8 It constitutes the middle buffer zone, and area T is the central reference area.

[0056] Furthermore, the dual-neighborhood block local contrast module is used to process the gradient mode feature map of the original infrared image to obtain the processed gradient mode feature map; the specific process is:

[0057] Slide the dual-neighborhood block local contrast module on the gradient modulus feature map, so that the center of the dual-neighborhood block local contrast module passes through each pixel point on the gradient modulus feature map in turn, obtain the processed gradient modulus feature of each pixel point, and use the processed gradient modulus features of each pixel point to form a processed gradient modulus feature map;

[0058] For a pixel point (x, y) on the gradient modulus feature map, when the center of the dual-neighborhood block local contrast module slides to the pixel point (x, y), the calculation method of the processed gradient modulus feature of the pixel point (x, y) is:

[0059] Step S31, calculating the dissimilarity between the reference area and each area in the middle buffer area and the dissimilarity between the reference area and each area in the outer background area;

[0060]

[0061] Among them, d(T,IB i ) represents the difference between the reference area and the ith area in the middle buffer zone; d(T,OB j ) represents the difference between the reference area and the jth area in the outer background area; m T Represents the mean value of the gradient modulus characteristics of each pixel in the reference area; m IBi represents the mean value of the gradient modulus feature of each pixel in the i-th region in the middle buffer, i = 1, 2, ..., 8; m OBj Represents the mean value of the gradient modulus feature of each pixel in the jth region in the outer background area, j = 1, 2, ..., 8;

[0062] Step S32, calculating the contrast difference between the reference area and the middle buffer area and the contrast difference between the reference area and the outer background area according to the result in step S31;

[0063] DTI=min[d(T,IB i )×d(T,IB i+4 )],i=1,2,3,4

[0064] DTO=min[d(T,OB j )×d(T,OB j+4 )],j=1,2,3,4

[0065] Among them, DTI is the contrast difference between the reference area and the middle buffer area, DTO is the contrast difference between the reference area and the outer background area, and “×” indicates multiplication operation;

[0066] Step S33, calculate the processed gradient modulus feature of the pixel point (x, y) according to DTI and DTO:

[0067] DNLMCM=DTI×DTO

[0068] Among them, DNLMCM is the processed gradient modulus feature of the pixel point (x, y).

[0069] Furthermore, the specific process of step S4 is as follows:

[0070] Take pixel (i, j) as an example

[0071]

[0072] Where f(i,j) represents the pixel value of pixel point (i,j); M×N is the size of the local window of the image centered at pixel point (i,j); ij is the grayscale distribution of pixel (i, j), H f is the entropy weight feature of pixel (i, j);

[0073] The entropy weight features of global pixels in the original infrared image form an entropy weight feature map.

[0074] Furthermore, the specific process of step S5 is as follows:

[0075]

[0076] in, represents pixel-by-pixel multiplication, GD′ is the azimuth feature map after filtering, DNLMCM′ is the gradient modulus feature map after processing, and H′ f is the entropy weight feature map, and FEGSDE is the final feature map.

[0077] The beneficial effects of the present invention are:

[0078] The gradient mode feature extracted by the present invention can make the target and other areas clearly distinguishable. By extracting the angular feature and combining it with filtering, the target area can be enhanced to achieve background area suppression. Taking the local information entropy as a weight factor can effectively supplement the information loss caused in the process of gradient domain feature extraction. Finally, by integrating the angular feature map after filtering, the gradient mode feature map after processing, and the entropy weight feature map, the accuracy of infrared small target detection can be improved. Since the method of the present invention has a very good detection effect on small targets in infrared images, the accuracy of small target detection and the universality of multiple usage environments are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 is an overall flow chart of the method of the present invention;

[0080] Figure 2 It is a schematic diagram of the direction angle filtering template;

[0081] Figure 3 It is a schematic diagram of the dual-neighborhood block local contrast module;

[0082] Figure 4 It is a sliding schematic diagram of the dual-neighborhood block local contrast module;

[0083] Figure 5 Schematic diagram of the real scene used in the experiment;

[0084] Figure 6 This is a schematic diagram of the detection effect under cloudy background conditions;

[0085] Figure 7 This is a schematic diagram of the detection effect under the condition of a few-cloud background;

[0086] Figure 8 This is a schematic diagram of the detection effect under surface background conditions;

[0087] Fig. 9 Schematic diagram of the detection effect under building background conditions. DETAILED DESCRIPTION

[0088] Specific implementation method 1: Combination Figure 1 The present embodiment describes a method for detecting small infrared targets based on grayscale difference and information entropy, and the method specifically includes the following steps:

[0089] Step S1, preprocessing the original infrared image based on the Scharr gradient operator, and obtaining a gradient modulus feature map and an angular feature map of the original infrared image according to the preprocessing result;

[0090] Step S2, constructing a direction angle filtering template, filtering the direction angle feature map through the direction angle filtering template, and extracting the filtered direction angle feature map of the original infrared image;

[0091] Step S3, constructing a dual-neighborhood block local contrast module, wherein the dual-neighborhood block local contrast module includes a central reference area, a middle buffer area and an outer background area;

[0092] The dual-neighborhood block local contrast module is used to process the gradient mode feature map of the original infrared image to obtain a processed gradient mode feature map;

[0093] Step S4, respectively calculating the entropy weight feature of each pixel in the original infrared image to obtain an entropy weight feature map of the original infrared image;

[0094] Step S5, integrating the filtered azimuth feature map, the processed gradient modulus feature map, and the entropy weight feature map of the original infrared image to obtain a final feature map of the original infrared image;

[0095] Step S6: Detect small targets in the original infrared image based on the obtained final feature map. The detection process can adopt existing methods.

[0096] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that the specific process of step S1 is as follows:

[0097] Step S11, using the Scharr gradient operator in the horizontal direction to perform convolution calculation on the original infrared image, respectively obtaining the horizontal gradient of each pixel point, and using the Scharr gradient operator in the vertical direction to perform convolution calculation on the original infrared image, respectively obtaining the vertical gradient of each pixel point;

[0098] Step S12: for any pixel point, calculate the gradient modulus feature and the direction angle feature of the pixel point according to the horizontal gradient and the vertical gradient of the pixel point;

[0099] Similarly, after traversing each pixel point in the original infrared image, the gradient modulus feature map and the directional angle feature map of the original infrared image are obtained.

[0100] The other steps and parameters are the same as those in the first embodiment.

[0101] Specific implementation method three: This implementation method is different from specific implementation method one or two in that the Scharr gradient operator in the horizontal direction and the Scharr gradient operator in the vertical direction are respectively:

[0102]

[0103] Among them, G x_new is the Scharr gradient operator in the horizontal direction;

[0104]

[0105] Among them, G y_new is the Scharr gradient operator in the vertical direction.

[0106] The other steps and parameters are the same as those in the first or second embodiment.

[0107] This implementation manner improves the existing Scharr operator. The improved Scharr operator can collect more image information when calculating the gradient by increasing the dimension.

[0108] Specific implementation method 4: This implementation method is different from any one of specific implementation methods 1 to 3 in that, in step S12, the calculation method of the gradient modulus feature and the directional angle feature of the pixel point is specifically as follows:

[0109] Taking the pixel point (x, y) as an example, the horizontal gradient of the pixel point (x, y) is recorded as The vertical gradient of the pixel point (x, y) is recorded as Then the gradient modulus feature of the pixel point (x, y) is:

[0110]

[0111] Among them, M(x,y) is the gradient modulus feature of the pixel point (x,y);

[0112] The direction angle feature of the pixel point (x, y) is:

[0113]

[0114] Among them, D(x,y) is the direction angle feature of the pixel point (x,y).

[0115] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0116] The target area has unique gradient characteristics, that is, the gradient direction angle of the target and its surrounding neighborhood points to the center, and the modulus of the gradient in the target area is larger; whether it is a complex background or a sharp geometric edge, its gradient characteristics are obviously different from the target area. Therefore, the gradient information of small targets can be used to help locate the edges and contours of the target, thereby better detecting and positioning.

[0117] Specific implementation method five: Combination Figure 2 Description of this embodiment. This embodiment differs from any one of the specific embodiments one to four in that the size of the directional angle filter template is 3×3, the elements in the first row and first column of the directional angle filter template are π / 4, the elements in the first row and second column are π / 2, the elements in the first row and third column are 3π / 4, the elements in the second row and first column are 0, the elements in the second row and second column are 0, the elements in the second row and third column are π, the elements in the third row and first column are -π / 4, the elements in the third row and second column are -π / 2, and the elements in the third row and third column are -3π / 4.

[0118] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0119] Specific implementation method 6: This implementation method is different from any one of specific implementation methods 1 to 5 in that, in step S2, the directional angle feature map is filtered by the directional angle filter template to extract the filtered directional angle feature map of the original infrared image. The specific process is as follows:

[0120] Slide the directional angle filter template on the directional angle feature map so that the center of the directional angle filter template passes through each pixel point on the directional angle feature map in sequence;

[0121] For the pixel point (x, y) on the angular feature map, the angular feature map composed of (x, y) and all the pixels in the 3×3 neighborhood of (x, y) is recorded as D′, and the angular feature map D′ is convolved with the angular filter template, and the convolution result is used as the filtered angular feature of the pixel point (x, y):

[0122] GD(x,y)=D′*DFT

[0123] Where * represents convolution operation, DFT represents azimuth filter template, and GD(x,y) is the azimuth feature of pixel point (x,y) after filtering.

[0124] The filtered angular features of all pixels constitute the filtered angular feature map of the original infrared image.

[0125] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5.

[0126] In order to fit the central directivity of the small target block, the angles in the directional angle filter template all point to the center, and the directional angle feature map is filtered using the directional angle filter template, so that the correlation response between the directional angle filter template and the feature to be processed is higher, and the calculated value will be larger, which can achieve target area enhancement; for pixels in other areas, their direction pointing has a low correlation with the directional angle filter template, and the calculated result is relatively low, thereby achieving background area suppression.

[0127] It should be noted that when the center of the directional angle filter template slides to the boundary pixel of the directional angle feature map, the directional angle filter template will exceed the boundary of the directional angle feature map. At this time, the directional angle feature corresponding to the position beyond the boundary in the template is considered to be 0.

[0128] Specific implementation method seven: Combination Figure 3 and Figure 4 This embodiment is different from any one of the first to sixth embodiments in that the size of the dual-neighborhood block local contrast module is 15×15; wherein:

[0129] The area consisting of pixels located in the first to third rows and the first to third columns in the dual-neighborhood block local contrast module is denoted as B. o1 ;

[0130] The area consisting of pixels located in the first to third rows and in the fourth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B o2 ;

[0131] The area consisting of pixels located in the first to third rows and the thirteenth to fifteenth columns in the dual-neighborhood block local contrast module is denoted as B. o3 ;

[0132] The area consisting of pixels located in the fourth to twelfth rows and the thirteenth to fifteenth columns in the dual-neighborhood block local contrast module is denoted as B o4 ;

[0133] The area consisting of pixels located in the 13th to 15th rows and the 13th to 15th columns in the dual-neighborhood block local contrast module is denoted as B. o5 ;

[0134] The area consisting of pixels located in the 13th to 15th rows and the 4th to 12th columns in the dual-neighborhood block local contrast module is denoted as B o6 ;

[0135] The area consisting of pixels located in the thirteenth to fifteenth rows and in the first to third columns in the dual-neighborhood block local contrast module is denoted as B o7 ;

[0136] The area consisting of pixels located in the fourth to twelfth rows and in the first to third columns in the dual-neighborhood block local contrast module is denoted as B o8 ;

[0137] The area consisting of pixels located in the fourth to sixth rows and the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i1 ;

[0138] The area consisting of pixels located in the fourth to sixth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as B i2 ;

[0139] The area consisting of pixels located in the fourth to sixth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i3 ;

[0140] The area consisting of pixels located in the seventh to ninth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i4 ;

[0141] The area consisting of pixels located in the tenth to twelfth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i5 ;

[0142] The area consisting of pixels located in the tenth to twelfth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as B. i6 ;

[0143] The area consisting of pixels located in the tenth to twelfth rows and in the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i7 ;

[0144] The area consisting of pixels located in the seventh to ninth rows and in the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i8 ;

[0145] The area consisting of pixels located in the seventh to ninth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as T;

[0146] Region B o1 , B o2 , B o3 , B o4 , B o5 , B o6 , B o7 and B o8 Composing the outer background area, area B i1 , B i2 , B i3 , B i4 , B i5 , B i6 , B i7 and B i8 It constitutes the middle buffer zone, and area T is the central reference area.

[0147] The other steps and parameters are the same as those in Specific Embodiments 1 to 6.

[0148] It should be noted that when the center of the module slides to the boundary pixel, that is, when the module exceeds the image boundary, the gradient modulus feature of the missing pixel is considered to be 0, that is, 0 is used to participate in the mean calculation of the gradient modulus in the region.

[0149] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that the dual-neighborhood block local contrast module is used to process the gradient mode feature map of the original infrared image to obtain the processed gradient mode feature map; the specific process is:

[0150] Slide the dual-neighborhood block local contrast module on the gradient modulus feature map, so that the center of the dual-neighborhood block local contrast module passes through each pixel point on the gradient modulus feature map in turn, obtain the processed gradient modulus feature of each pixel point, and use the processed gradient modulus features of each pixel point to form a processed gradient modulus feature map;

[0151] For a pixel point (x, y) on the gradient modulus feature map, when the center of the dual-neighborhood block local contrast module slides to the pixel point (x, y), the calculation method of the processed gradient modulus feature of the pixel point (x, y) is:

[0152] Step S31, calculating the dissimilarity (i.e., grayscale difference) between the reference area and each area in the middle buffer area, and the dissimilarity between the reference area and each area in the outer background area;

[0153]

[0154] Among them, d(T,IB i ) represents the difference between the reference area and the ith area in the middle buffer zone; d(T,OB j ) represents the difference between the reference area and the jth area in the outer background area; m T Represents the mean value of the gradient modulus characteristics of each pixel in the reference area; m IBi represents the mean value of the gradient modulus feature of each pixel in the i-th region in the middle buffer, i = 1, 2, ..., 8; m OBj Represents the mean value of the gradient modulus feature of each pixel in the jth region in the outer background area, j = 1, 2, ..., 8;

[0155] Step S32, calculating the contrast difference between the reference area and the middle buffer area and the contrast difference between the reference area and the outer background area according to the result in step S31;

[0156] DTI=min[d(T,IB i )×d(T,IB i+4 )],i=1,2,3,4

[0157] DTO=min[d(T,OB j )×d(T,OB j+4 )],j=1,2,3,4

[0158] Among them, DTI is the contrast difference between the reference area and the middle buffer area, DTO is the contrast difference between the reference area and the outer background area, and “×” indicates multiplication operation;

[0159] Step S33, calculate the processed gradient modulus feature of the pixel point (x, y) according to DTI and DTO:

[0160] DNLMCM=DTI×DTO

[0161] Among them, DNLMCM is the processed gradient modulus feature of the pixel point (x, y).

[0162] The other steps and parameters are the same as those in Specific Embodiments 1 to 7.

[0163] In order to maximize the contrast between the target area and the background area and improve the characteristic performance of the target area, the present invention selects the minimum difference between the reference block and the surrounding background blocks as the contrast measure. The characteristic saliency of small targets is enhanced by double-layer contrast product, and the gradient modulus contrast information of the center and the neighborhood layers is effectively utilized to effectively suppress the background area and the strong edge area, so that the characteristic information of the target area is retained and enhanced. At the same time, the gradient modulus mean is used as the final judgment object, which can effectively reduce the impact of a single highlight background or noise.

[0164] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that the specific process of step S4 is as follows:

[0165] Take pixel (i, j) as an example

[0166]

[0167] Where f(i,j) represents the pixel value of pixel point (i,j), and its value range is [0,255]; M×N is the size of the local window of the image centered at pixel point (i,j), and p ij is the grayscale distribution of pixel (i, j), H f is the entropy weight feature of the pixel point (i, j), and the base of the logarithm in the present invention may be 10;

[0168] The entropy weight features of global pixels in the original infrared image form an entropy weight feature map.

[0169] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0170] This embodiment uses the local information entropy of each pixel in the original infrared image as a weight factor. By comparing the local information entropy of different areas, it is possible to determine which areas contain more information and complexity, thereby giving them higher weights. In addition, by performing entropy calculation on the original infrared image, it is also possible to effectively supplement the information loss caused in the gradient domain feature extraction process.

[0171] It should be noted that when the center of the local window slides to the boundary pixel, that is, when the local window exceeds the image boundary, the pixel value of the missing pixel is considered to be 0, that is, 0 is used to participate in the calculation.

[0172] Specific implementation method 10: This implementation method is different from the specific implementation methods 1 to 9 in that the specific process of step S5 is as follows:

[0173]

[0174] in, represents pixel-by-pixel multiplication, GD′ is the azimuth feature map after filtering, DNLMCM′ is the gradient modulus feature map after processing, and H′ f is the entropy weight feature map, and FEGSDE is the final feature map.

[0175] The other steps and parameters are the same as those in Specific Embodiments 1 to 9.

[0176] This embodiment integrates the filtered azimuth feature map, the processed gradient modulus feature map, and the entropy weight feature map to form a comprehensive feature representation. This integration method not only considers the local structural information of the image, but also takes into account the local statistical characteristics of the image, thereby effectively improving the accuracy of infrared small target feature extraction.

[0177] Experimental Section

[0178] The method of the present invention is verified by specific experiments. In order to evaluate the method of the present invention, combined with the application scenario of the small target detection device, different application background images are manually selected in the general infrared small target data set SIRST, and four environments of cloudy, light cloud, surface, and building are distinguished. The initial image and the corresponding label mask in the data set are reintegrated to establish four sets of infrared sequence experimental image sample sets as the final experimental verification data. The typical sample examples of each sequence are as follows: Figure 5 As shown in the figure, the target area is selected and a local magnified image is given. The information of the four sets of sequence images is shown in Table 1:

[0179] Table 1 Sequence image information table

[0180]

[0181] In the comparative experiment, we use seven feature extraction baseline methods based on physical properties as comparative references to evaluate the FEGSDE algorithm proposed in this invention. The comparative methods include TopHat algorithm, MaxMean algorithm, MGDWE algorithm, MPCM algorithm, LIG ​​algorithm, ADMD algorithm and AADCDD algorithm. Different methods have different parameter settings. The hyperparameter settings of the method proposed in this invention and the comparative baseline method are shown in Table 2.

[0182] Table 2 Algorithm hyperparameter settings

[0183]

[0184] The detection effects of the method of the present invention and other comparison methods under cloudy, lightly cloudy, ground surface and building background conditions are as follows: Figure 6 , Figure 7 , Figure 8 and Fig. 9 shown.

[0185] A quantitative analysis was performed to verify the feasibility of the algorithm of the present invention. In order to give the detection results of the algorithm more objectively, Table 3 shows the quantitative detection evaluation results of eight feature extraction algorithms under four different sequence backgrounds, including the background suppression factor BSF value and the signal-to-noise ratio gain SCRG, where the bold part is the maximum value.

[0186] Table 3 Comparison of results of different algorithms

[0187]

[0188] By comparing the BSF values ​​of the algorithms under different backgrounds, it is found that the BSF obtained under light cloud conditions is higher, while that under cloudy and ground backgrounds is slightly lower. This can explain that for different small target detection algorithms, cloud clutter and surface vegetation interference test the background suppression ability of the algorithm more. By selecting a single scene for longitudinal comparison, it is found that the BSF values ​​of the TopHat algorithm, MaxMean algorithm, and MPCM algorithm based on a single feature are significantly lower than those of the fusion algorithm based on multiple physical features. This can indirectly show that from the perspective of multiple features, achieving complementary advantages between features has a certain impact on improving the algorithm capabilities.

[0189] Taking any sequence background, the SCRG values ​​of each algorithm are compared vertically. It is found that the single feature algorithm has a lower SCRG value than the multi-feature algorithm, indicating that the multi-feature fusion algorithm has a better target enhancement effect than the single feature algorithm. Overall, the FEGSDE algorithm of the present invention can achieve good SCRG under different sequence backgrounds, indicating that the algorithm of the present invention has a strong target feature enhancement ability.

[0190] In general, the FEGSDE algorithm of the present invention can achieve significantly better results in four different sequence backgrounds. Only in the cloudy background, the average BSF value of the FEGSDE algorithm of the present invention is not significantly ahead of other algorithms, while in the other three backgrounds, the BSF value of the FEGSDE algorithm of the present invention is significantly large, which can show that the FEGSDE algorithm proposed in the present invention has a strong suppression ability for different background environments and can achieve a higher suppression effect on various background clutter interferences.

[0191] The above calculation examples of the present invention are only used to explain the calculation model and calculation process of the present invention in detail, and are not intended to limit the implementation methods of the present invention. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is impossible to list all the implementation methods here. All obvious changes or modifications derived from the technical solution of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for detecting small infrared targets based on grayscale difference and information entropy, characterized in that: The method specifically comprises the following steps: Step S1, preprocessing the original infrared image based on the Scharr gradient operator, and obtaining a gradient modulus feature map and an angular feature map of the original infrared image according to the preprocessing result; Step S2, constructing a direction angle filtering template, filtering the direction angle feature map through the direction angle filtering template, and extracting the filtered direction angle feature map of the original infrared image; Step S3, constructing a dual-neighborhood block local contrast module, wherein the dual-neighborhood block local contrast module includes a central reference area, a middle buffer area and an outer background area; The dual-neighborhood block local contrast module is used to process the gradient mode feature map of the original infrared image to obtain a processed gradient mode feature map; Step S4, respectively calculating the entropy weight feature of each pixel in the original infrared image to obtain an entropy weight feature map of the original infrared image; Step S5, integrating the filtered azimuth feature map, the processed gradient modulus feature map, and the entropy weight feature map of the original infrared image to obtain a final feature map of the original infrared image; Step S6: Detect small targets in the original infrared image based on the obtained final feature map.

2. The infrared small target detection method based on grayscale difference and information entropy according to claim 1 is characterized in that: The specific process of step S1 is as follows: Step S11, using the Scharr gradient operator in the horizontal direction to perform convolution calculation on the original infrared image, respectively obtaining the horizontal gradient of each pixel point, and using the Scharr gradient operator in the vertical direction to perform convolution calculation on the original infrared image, respectively obtaining the vertical gradient of each pixel point; Step S12: for any pixel point, calculate the gradient modulus feature and the direction angle feature of the pixel point according to the horizontal gradient and the vertical gradient of the pixel point; Similarly, after traversing each pixel point in the original infrared image, the gradient modulus feature map and the directional angle feature map of the original infrared image are obtained.

3. The infrared small target detection method based on grayscale difference and information entropy according to claim 2 is characterized in that: The Scharr gradient operator in the horizontal direction and the Scharr gradient operator in the vertical direction are respectively: Among them, G x_new is the Scharr gradient operator in the horizontal direction; Among them, G y_new is the Scharr gradient operator in the vertical direction.

4. The infrared small target detection method based on grayscale difference and information entropy according to claim 3 is characterized in that: In step S12, the calculation method of the gradient modulus feature and the directional angle feature of the pixel point is specifically as follows: Taking the pixel point (x, y) as an example, the horizontal gradient of the pixel point (x, y) is recorded as The vertical gradient of the pixel point (x, y) is recorded as Then the gradient modulus feature of the pixel point (x, y) is: Among them, M(x,y) is the gradient modulus feature of the pixel point (x,y); The direction angle feature of the pixel point (x, y) is: Among them, D(x,y) is the direction angle feature of the pixel point (x,y).

5. The infrared small target detection method based on grayscale difference and information entropy according to claim 4 is characterized in that: The size of the directional angle filter template is 3×3, the elements of the first row and first column in the directional angle filter template are π / 4, the elements of the first row and second column are π / 2, the elements of the first row and third column are 3π / 4, the elements of the second row and first column are 0, the elements of the second row and second column are 0, the elements of the second row and third column are π, the elements of the third row and first column are -π / 4, the elements of the third row and second column are -π / 2, and the elements of the third row and third column are -3π / 4.

6. The infrared small target detection method based on grayscale difference and information entropy according to claim 5 is characterized in that: In step S2, the directional angle feature map is filtered by the directional angle filter template to extract the filtered directional angle feature map of the original infrared image. The specific process is as follows: Slide the directional angle filter template on the directional angle feature map so that the center of the directional angle filter template passes through each pixel point on the directional angle feature map in sequence; For the pixel point (x, y) on the angular feature map, the angular feature map composed of (x, y) and all the pixels in the 3×3 neighborhood of (x, y) is recorded as D′, and the angular feature map D′ is convolved with the angular filter template, and the convolution result is used as the filtered angular feature of the pixel point (x, y): GD(x,y)=D′*DFT Where * represents convolution operation, DFT represents directional filter template, and GD(x,y) is the directional feature of the pixel point (x,y) after filtering. The filtered angular features of all pixels constitute the filtered angular feature map of the original infrared image.

7. The infrared small target detection method based on grayscale difference and information entropy according to claim 6 is characterized in that: The size of the dual-neighborhood block local contrast module is 15×15; wherein: The area consisting of pixels located in the first to third rows and the first to third columns in the dual-neighborhood block local contrast module is denoted as B. o1 ; The area consisting of pixels located in the first to third rows and in the fourth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B o2 ; The area consisting of pixels located in the first to third rows and the thirteenth to fifteenth columns in the dual-neighborhood block local contrast module is denoted as B. o3 ; The area consisting of pixels located in the fourth to twelfth rows and the thirteenth to fifteenth columns in the dual-neighborhood block local contrast module is denoted as B o4 ; The area consisting of pixels located in the 13th to 15th rows and the 13th to 15th columns in the dual-neighborhood block local contrast module is denoted as B. o5 ; The area consisting of pixels located in the 13th to 15th rows and the 4th to 12th columns in the dual-neighborhood block local contrast module is denoted as B o6 ; The area consisting of pixels located in the thirteenth to fifteenth rows and in the first to third columns in the dual-neighborhood block local contrast module is denoted as B o7 ; The area consisting of pixels located in the fourth to twelfth rows and in the first to third columns in the dual-neighborhood block local contrast module is denoted as B o8 ; The area consisting of pixels located in the fourth to sixth rows and the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i1 ; The area consisting of pixels located in the fourth to sixth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as B i2 ; The area consisting of pixels located in the fourth to sixth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i3 ; The area consisting of pixels located in the seventh to ninth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i4 ; The area consisting of pixels located in the tenth to twelfth rows and in the tenth to twelfth columns in the dual-neighborhood block local contrast module is denoted as B i5 ; The area consisting of pixels located in the tenth to twelfth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as B. i6 ; The area consisting of pixels located in the tenth to twelfth rows and in the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i7 ; The area consisting of pixels located in the seventh to ninth rows and in the fourth to sixth columns in the dual-neighborhood block local contrast module is denoted as B i8 ; The area consisting of pixels located in the seventh to ninth rows and the seventh to ninth columns in the dual-neighborhood block local contrast module is denoted as T; Region B o1 , B o2 , B o3 , B o4 , B o5 , B o6 , B o7 and B o8 Composing the outer background area, area B i1 , B i2 , B i3 , B i4 , B i5 , B i6 , B i7 and B i8 It constitutes the middle buffer zone, and area T is the central reference area.

8. The infrared small target detection method based on grayscale difference and information entropy according to claim 7 is characterized in that: The dual-neighborhood block local contrast module is used to process the gradient mode feature map of the original infrared image to obtain the processed gradient mode feature map; the specific process is: Slide the dual-neighborhood block local contrast module on the gradient modulus feature map, so that the center of the dual-neighborhood block local contrast module passes through each pixel point on the gradient modulus feature map in turn, obtain the processed gradient modulus feature of each pixel point, and use the processed gradient modulus features of each pixel point to form a processed gradient modulus feature map; For a pixel point (x, y) on the gradient modulus feature map, when the center of the dual-neighborhood block local contrast module slides to the pixel point (x, y), the calculation method of the processed gradient modulus feature of the pixel point (x, y) is: Step S31, calculating the dissimilarity between the reference area and each area in the middle buffer area and the dissimilarity between the reference area and each area in the outer background area; Among them, d(T,IB i ) represents the difference between the reference area and the ith area in the middle buffer zone; d(T,OB j ) represents the difference between the reference area and the jth area in the outer background area; m T Represents the mean value of the gradient modulus characteristics of each pixel in the reference area; m IBi represents the mean value of the gradient modulus feature of each pixel in the i-th region in the middle buffer, i = 1, 2, ..., 8; m OBj Represents the mean value of the gradient modulus feature of each pixel in the jth region in the outer background area, j = 1, 2, ..., 8; Step S32, calculating the contrast difference between the reference area and the middle buffer area and the contrast difference between the reference area and the outer background area according to the result in step S31; <h2 style=";text-align:left;direction:ltr">DTI = min[d(T,IB<h2 style=";text-align:left;direction:ltr"> i <h2 style=";text-align:left;direction:ltr"> )×d(T,IB<h2 style=";text-align:left;direction:ltr"> i+4 <h2 style=";text-align:left;direction:ltr"> )],i = 1, 2, 3, 4 DTO=min[d(T,OB j )×d(T,OB j+4 )],j=1,2,3,4 Among them, DTI is the contrast difference between the reference area and the middle buffer area, DTO is the contrast difference between the reference area and the outer background area, and "×" represents multiplication operation; Step S33, calculate the processed gradient modulus feature of the pixel point (x, y) according to DTI and DTO: DNLMCM=DTI×DTO Among them, DNLMCM is the processed gradient modulus feature of the pixel point (x, y).

9. The infrared small target detection method based on grayscale difference and information entropy according to claim 8, characterized in that: The specific process of step S4 is as follows: Take pixel (i, j) as an example Where f(i,j) represents the pixel value of pixel point (i,j); M×N is the size of the local window of the image centered at pixel point (i,j); ij is the grayscale distribution of pixel (i, j), H f is the entropy weight feature of pixel (i, j); The entropy weight features of global pixels in the original infrared image form an entropy weight feature map.

10. The infrared small target detection method based on grayscale difference and information entropy according to claim 9, characterized in that: The specific process of step S5 is as follows: in, represents pixel-by-pixel multiplication, GD′ is the azimuth feature map after filtering, DNLMCM′ is the gradient modulus feature map after processing, and H′ f is the entropy weight feature map, and FEGSDE is the final feature map.

Images