CN115170523B

CN115170523B - Low-complexity infrared dim target detection method based on local contrast

Info

Publication number: CN115170523B
Application number: CN202210833296.XA
Authority: CN
Inventors: 马永奎; 王俊杰; 张佳岩; 赵洪林; 单成兆
Original assignee: Harbin Institute of Technology
Current assignee: Harbin Institute of Technology
Priority date: 2022-07-14
Filing date: 2022-07-14
Publication date: 2023-05-05
Anticipated expiration: 2042-07-14
Also published as: CN115170523A

Abstract

A low-complexity infrared weak and small target detection method based on local contrast relates to the technical field of target detection, and aims at solving the problem of low target detection speed under a complex background in the prior art, the method utilizes the characteristics that the infrared weak and small target has larger difference from the background and has higher contrast locally to detect the infrared weak and small target, and the transmission bandwidth of a digital system is fully utilized through an efficient filtering template in the implementation process, so that the image traversing time is reduced, and the system processing speed is greatly improved; meanwhile, the algorithm complexity is reduced, floating point operation and nonlinear operation are reduced, the processing time delay of the system is reduced, the design of a large-scale assembly line is facilitated, and the hardware implementation difficulty is reduced. Finally, the method can realize reliable high-speed infrared weak and small target detection, the throughput of the actually measured system reaches 20Gbps, the detection rate is higher than 90%, the false alarm rate is lower than 10%, and the high-speed detection can be realized in engineering application.

Description

Low-complexity infrared dim target detection method based on local contrast

Technical Field

The invention relates to the technical field of target detection, in particular to a low-complexity infrared dim target detection method based on local contrast.

Background

The infrared imaging target detection system has the advantages of capability of working day and night, no interference from natural environments such as cloud and fog, and the like, and strong anti-interference capability. Therefore, the infrared imaging system is widely applied to the military, industrial and civil fields of all-weather security monitoring, unmanned aerial vehicle disaster relief detection and weapon guidance.

Low signal-to-noise ratio small object detection in complex contexts is a very challenging task because objects in an image are typically very small and lack specific shape, texture and structure information, which results in small objects often occupying very small pixels in an image and being strongly disturbed by background clutter or random noise. In order to solve the above problems, the mechanism of the human visual system has been applied in the field of infrared dim target detection in recent years, which mainly uses the visual attention mechanism, the brightness and contrast sensitivity mechanism of human eyes, and the like, and adopts saliency maps to realize feature extraction. The detection algorithm based on human eye vision is mainly divided into 2 types, the first type is an algorithm based on spectrum residual error, the main principle is that the spectrum of an infrared image is subjected to a low-pass filter to obtain a smoothed spectrum, then the smoothed spectrum is subjected to difference with the original spectrum to obtain spectrum residual error, and the small target position is detected based on high-frequency components in the residual error. The other type is a detection algorithm based on local contrast, which uses the big difference between the brightness of a small target and the neighborhood of the small target, and adopts the idea of a sliding window to detect the contrast in the sliding window and search the position of the small target.

Because the local contrast measurement algorithm has small calculation amount and good detection effect compared with a frequency spectrum residual error method, the method is used and improved by a plurality of scholars, the improvement thinking is to improve the mode of calculating the contrast and to improve the window, but the algorithm is designed aiming at software implementation, and a plurality of floating points and nonlinear operations are used, so that the method is not beneficial to parallel implementation of a high-speed digital image processing system based on an FPGA; due to the serial processing capability of the CPU or the DSP, the algorithm has high processing time delay and poor instantaneity, and meanwhile, the filtering window is not selected too large, so that the time for traversing a pair of images is too long, and the limited bandwidth resource in the transmission process of the digital signal is not fully utilized.

Disclosure of Invention

The purpose of the invention is that: aiming at the problem of low target detection speed under a complex background in the prior art, the low-complexity infrared weak and small target detection method based on local contrast is provided.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a low-complexity infrared dim target detection method based on local contrast comprises the following steps:

step one: acquiring an infrared image of an object to be identified, and preprocessing the infrared image of the object to be identified, wherein the preprocessing comprises morphological open operation filtering processing and maximum median filtering processing;

step two: the method comprises the steps of carrying out contrast detection on a preprocessed infrared image by using a high-efficiency filtering template, wherein the high-efficiency filtering template is a square template, and the high-efficiency filtering template is 2 ⁿ ×2 ⁿ The subblocks are formed, n is more than or equal to 2 and less than or equal to 5, and the size of each subblock is 8 multiplied by 8 pixels;

the contrast detection comprises the following steps:

step two,: utilizing 3 multiplied by 3 sub-blocks in the efficient filtering template as a processing unit, wherein the central sub-block of the processing unit is a target area to be processed, and the rest sub-blocks are background areas;

step two: the processing unit is used as a measuring window to process the preprocessed infrared image, and the specific steps are as follows:

step two, one by one: sequencing all pixels in a target area to be processed by gray values, taking the maximum value and the next maximum value of the gray values, and calculating the average value of the maximum value and the next maximum value, wherein the average value is used as an A value;

step two, two: calculating the average value of gray values of pixels in the sub-blocks corresponding to the background area, wherein the average value is taken as a B value;

step two, one and three: acquiring three sub-blocks corresponding to one direction of 8 directions of a target area to be processed, and then acquiring the average value of B values of the three sub-blocks, wherein the value is taken as a C value;

step two, one and four: the value A and the value C are subjected to difference, and the result is the contrast measurement value of the corresponding direction in 8 directions of the target area to be processed;

step two and step five: repeating the second step, the third step and the fourth step to obtain contrast measurement values in 8 directions;

the target area in the processing unit is area ₀ The background area is area from left to right and from top to bottom ₁ 、area ₂ 、area ₃ 、area ₄ 、area ₅ 、area ₆ 、area ₇ 、area ₈ ；

The 8 directions are specifically as follows: up direction neighborhood (area) ₁ ,area ₂ ,area ₃ ) Down direction neighborhood (area) ₆ ,area ₇ ,area ₈ ) Left direction neighborhood (area) ₁ ,area ₄ ,area ₆ ) Right direction neighborhood (area) ₃ ,area ₅ ,area ₈ ) A first quadrant neighborhood (area) ₂ ,area ₃ ,area ₅ ) A second quadrant neighborhood (area) ₂ ,area ₁ ,area ₄ ) A third quadrant neighborhood (area) ₄ ,area ₆ ,area ₇ ) A fourth quadrant neighborhood (area) ₅ ,area ₈ ,area ₇ )；

Step three: taking the target area to be processed as a pixel of the saliency map, taking absolute values of contrast measurement values in 8 directions, and summing the absolute values to obtain a gray value of the pixel;

step four: repeating the second step and the third step until the infrared image is processed, so as to obtain a saliency map;

step five: and (3) carrying out threshold segmentation on the saliency map by using the constant false alarm rate, extracting the region with the contrast larger than the threshold value from the segmentation result, and carrying out position output.

Further, the efficient filtering template is a double-layer template, the efficient filtering template comprises a first-layer template and a second-layer template, and the first-layer template is 2 ⁿ ×2 ⁿ Square templates of sub-blocks, the second layer template is 2 ⁿ -1×2 ⁿ -a square template of 1 sub-block, every four adjacent sub-blocks in the first layer template corresponding to one sub-block in the second layer template, the intersection points of the four adjacent sub-blocks in the first layer template coinciding with the center points of the sub-blocks in the second layer template.

Further, the specific steps of contrast detection are as follows:

step 1: area of the target area ₀ All pixels in the target region are ordered according to gray values, and the average value of the maximum pixel and the next maximum pixel of the gray values is used as the saliency brightness Lmax of the target region;

step 2: carrying out average pooling operation on each background area to obtain the average value of the gray level of each background area, and using the average value of the gray level of each background area to represent the saliency brightness Lmean of the background area _i ，i＝1,2,…,8；

Step 3: saliency Lmean according to background region _i Obtaining the saliency Lneighbor of each direction _m ：

Step 4: and obtaining a contrast measurement value by utilizing the difference between the saliency brightness of the target area and the saliency brightness of the target area in 8 directions.

Further, the saliency luminance Lmax of the target area is expressed as:

/>

wherein G is ¹ 、G ² Representing the maximum value pixel and the next-maximum value pixel of the target area.

Further, the salient luminance Lmean of the background area _i Expressed as:

wherein G is ⁱ (s, t) represents the gray value at the position (s, t) of the i-th block region, i being the number of the 3 sub-blocks required for the contrast in the current direction.

Further, the significance luminance Lneighbor of each direction _m Expressed as:

further, the contrast measurement is expressed as:

contrast _m ＝Lmax-Lneighbor _m

where m=1, 2, …,8.

Further, the threshold is expressed as:

T _h ＝m _c +k·σ

wherein m is _c The background clutter mean sigma is the root mean square of noise.

Further, the infrared image is acquired by an infrared camera.

The beneficial effects of the invention are as follows:

the infrared weak and small target detection method utilizes the characteristic that the infrared weak and small target has larger difference with the background and has higher contrast in local to carry out infrared weak and small target detection, and the transmission bandwidth of the digital system is fully utilized through the efficient filtering template in the implementation process, so that the image traversing time is reduced, and the system processing speed is greatly improved; meanwhile, the algorithm complexity is reduced, floating point operation and nonlinear operation are reduced, the processing time delay of the system is reduced, the design of a large-scale assembly line is facilitated, and the hardware implementation difficulty is reduced. Finally, the method can realize reliable high-speed infrared weak and small target detection, the throughput of the actually measured system reaches 20Gbps, the detection rate is higher than 90%, the false alarm rate is lower than 10%, and the high-speed detection can be realized in engineering application.

The method utilizes the characteristic that the infrared dim target has larger difference with the background and has higher contrast locally to detect the infrared dim target, a novel efficient filtering template is also used in detection, the efficient filtering template is a square template with the size of 2 and the even power number of pixels, and the divided sub-blocks in the template are squares with the size of 2 and the power number of pixels, and as the digital system generally takes the power of 2 as the transmission bit width, the method can be matched with the transmission bandwidth of various digital systems. The transmission bandwidth of the digital system is fully utilized through the efficient filtering template, so that the image traversing time can be reduced, and the system processing speed is greatly improved; meanwhile, the algorithm complexity is reduced, floating point operation and nonlinear operation are reduced, the processing time delay of the system is reduced, the design of a large-scale assembly line is facilitated, and the hardware implementation difficulty is reduced. Finally, reliable and rapid detection of the infrared weak target is realized.

Drawings

FIG. 1 is a flow chart of the present application;

FIG. 2 is a first layer schematic of a high efficiency filtering template;

FIG. 3 is a second layer schematic of the high efficiency filter template;

FIG. 4 is a schematic diagram of a measurement cell;

FIG. 5 is a schematic view of 8 measurement contrast directions;

FIG. 6 is a block diagram of a system hardware platform;

FIG. 7 is a schematic diagram of an embodiment of the present application;

FIG. 8 is a schematic diagram of an original infrared image in infrared small target detection;

FIG. 9 is a schematic view of a salient image obtained after processing in infrared dim target detection;

FIG. 10 is a schematic diagram of the detection result in infrared dim target detection;

FIG. 11 is a schematic diagram of a generic filtering template 1;

FIG. 12 is a schematic diagram of a generic filtering template 2;

FIG. 13 is a schematic diagram of a generic filtering template 3;

FIG. 14 is a schematic diagram of a high efficiency filtering template usage process;

FIG. 15 is a diagram of the actual correspondence of the high-efficiency filtering templates;

FIG. 16 is a schematic diagram of a relationship between a transmission unit and a processing unit;

FIG. 17 is a schematic diagram of a saliency map generation process 1;

FIG. 18 is a schematic diagram of a saliency map generation process 2;

FIG. 19 is a schematic diagram of a saliency map generation process 3;

FIG. 20 is an original view;

FIG. 21 is a saliency map;

fig. 22 is a schematic diagram of correspondence of the saliency map.

Detailed Description

It should be noted in particular that, without conflict, the various embodiments disclosed herein may be combined with each other.

The first embodiment is as follows: referring to fig. 1, the method for detecting a low-complexity infrared weak and small target based on local contrast according to the present embodiment specifically includes the following steps:

step 1: preprocessing an infrared image;

step 1.1: morphological open operation filtering: morphological opening operations are typically used to smooth the contours of objects, eliminating elongated protrusions. Highlight noise in the preprocessed image will be filtered out. The morphological open operation process is as follows:

the formula represents that the structural element B performs the following operation on the image A: first B corrodes a, then B swells a.

Step 1.2: maximum median filtering: since small objects are easily submerged by background clutter of an infrared image, the infrared background can be predicted by utilizing the regional correlation of the background clutter, and the object or noise and the background have no correlation and cannot be predicted, so that after the predicted background is subtracted from the original image, only a residual image containing the object and the noise is left. The method is realized by adopting a maximum median filtering method, the gray value of each pixel point is set to be the maximum value of the median values along different directions in a window of a certain field of the point, and the detail characteristics of a small target can be well protected. The algorithm process is to take a sliding window on the image, calculate the median value on each line along the horizontal, vertical and two diagonal directions respectively, compare the four median values, and then take the original value at the position of the largest substituted center pixel. The maximum median filter formula is:

y(x,y) _Max-median ＝max(z ₁ ,z ₂ ,z ₃ ,z ₄ ) (2)

z in formula (VI) _i Representing the condition of a certain dotted lineThe median of the dry pixels corresponds to the mathematical expression:

z ₁ ＝median{x(m,n-N),...,x(m,n),...,x(m,n+N)} (3)

z ₂ ＝median{x(m-N,n),...,x(m,n),...,x(m-N,n)} (4)

z ₃ ＝median{x(m+N,n-N),...,x(m,n),...,x(m-N,n+N)} (5)

z ₄ ＝median{x(m-N,n-N),...,x(m,n),...,x(m+N,n+N)} (6)

the method can effectively inhibit the background, enhance the contrast between the target and the surrounding background, and facilitate the subsequent processing. The residual image which only contains the target and the noise after the background inhibition can be obtained by making the difference between the background predicted by the method and the original image.

Step 2: multidirectional contrast detection:

the application utilizes the high-efficiency filtering template to carry out contrast detection on the preprocessed infrared image, the high-efficiency filtering template is a square template, and the high-efficiency filtering template comprises 2 ⁿ ×2 ⁿ Sub-blocks, n is more than or equal to 2 and less than or equal to 5, and the size of each sub-block is 8 multiplied by 8 pixels; processing the infrared image by taking 3 multiplied by 3 subblocks in the efficient filtering template as a processing unit, wherein the central subblock of the processing unit is a target area to be processed, and the rest subblocks are background areas;

for example: the efficient filtering template comprises 4×4 sub-blocks, and the processing unit is 3×3 sub-blocks, as shown in fig. 2, where the target area to be processed is four sub-blocks of "5, 6, 9, 10";

taking a sub-block '10' as an example, firstly sequencing all pixels in the sub-block '10' by gray values, then taking the maximum value and the next maximum value of the gray values, and then calculating the average value of the maximum value and the next maximum value;

then calculating the mean value of the pixel values of three sub-blocks in each of the 8 directions of the sub-block '10', namely the mean value of '13, 14 and 15', the mean value of '5, 6 and 7', the mean value of '7, 11 and 15', the mean value of '5, 9 and 13', the mean value of '9, 13 and 14', the mean value of '14, 15 and 11', the mean value of '11, 7 and 6', and the mean value of '9, 5 and 6';

respectively differencing the average value of the maximum value and the next maximum value with the average value of the pixel values of the three sub-blocks in each of the 8 directions to obtain 8 contrast measurement values;

the saliency map is a compressed sub-block, the sub-block corresponding to the target area to be processed is compressed to be used as a pixel of the saliency map, and the pixel value of the pixel is 8 contrast measurement values, takes absolute values and sums the absolute values.

In addition, the high efficiency filtering template may be 2 layers, the first layer comprising 2 ⁿ ×2 ⁿ Sub-blocks, the second layer comprising 2 ⁿ -1×2 ⁿ -1 sub-block, the first layer and the second layer being square templates, each four adjacent sub-blocks in the first layer template corresponding to one sub-block in the second layer template, the intersection points of the four adjacent sub-blocks in the first layer template coinciding with the center points of the sub-blocks in the second layer template;

in application, for example, the first layer comprises 4×4 sub-blocks, the second layer comprises 3×3 sub-blocks, the first layer template and the second layer template are parallelized, the result of the first layer template is used as the first layer of the saliency map, the result of the second layer template is used as the second layer of the saliency map, the two layers of efficient filtering templates can obtain 5 pixel points of the saliency map at a time, and the one layer of efficient filtering templates can only obtain 4 pixel points at a time. The order of the sliding of the sub-blocks is from top to bottom and from left to right.

area ₀ In order to enhance the contrast of the target, the present application sorts all pixels of the target region, and takes out the average value of the maximum value and the next-largest value to represent the significant brightness of the region, where the definition of the brightness as Lmax has a calculation formula:

g in ¹ 、G ² Representing the maximum value pixel and the next-maximum value pixel of the target area.

area ₁ ～area ₈ The area is a reference background area, the inventionThe mean-pooling operation is performed on each region, and the mean value is used to represent the salient brightness of the region, and the brightness is recorded as Lmean _i Then there is a calculation formula:

g in ⁱ (s, t) represents the gradation value at the position (s, t) of the i-th block region.

After having the salient luminance of each region, the present application continues to define the neighborhood luminance in eight directions. First define a neighborhood (area ₁ ,area ₂ ,area ₃ ) For the upward neighborhood, a similar definition (area ₆ ,area ₇ ,area ₈ ) Is a down direction neighborhood (area) ₁ ,area ₄ ,area ₆ ) Is left direction neighborhood (area) ₃ ,area ₅ ,area ₈ ) Is a right direction neighborhood (area) ₂ ,area ₃ ,area ₅ ) Is a first quadrant neighborhood (area) ₂ ,area ₁ ,area ₄ ) Is a second quadrant neighborhood (area) ₄ ,area ₆ ,area ₇ ) Is a third quadrant neighborhood (area) ₅ ,area ₈ ,area ₇ ) Is a fourth quadrant neighborhood. Defining each neighborhood brightness as Lneighbor _i Then there is a calculation formula:

lmean in _k Representing the saliency brightness of the three neighborhoods in the corresponding direction, m=1, 2, …, representing 8 directions, i representing the number of 3 sub-blocks needed to calculate the contrast in the current direction.

After defining the saliency luminance in the eight-direction neighborhood, the present application represents the contrast in a certain direction by using the difference between the saliency luminance of the target area and the saliency luminance in that direction _i Namely the formula:

contrast _m ＝Lmax-Lneighbor _m (10)

when the target sliding window traverses the whole picture, the contrast of the infrared image in 8 directions in a certain area can be obtained.

Step 3: constant false alarm rate threshold segmentation: the constant false alarm rate characteristic can keep the false alarm rate of the infrared detection system unchanged. Infrared image noise can be generally described by a gaussian probability density function with zero mean:

m is in _c Is the mean value of background clutter; m is m _t The real strength of the target point; sigma is the root mean square of noise; x is x _i Is the gray value somewhere in the image.

Further signal-to-noise ratio and threshold signal-to-noise ratio of the signal may be defined:

t in _h Representing a threshold segmentation threshold, the system considers that an object is detected when the gray value at a certain place in the image is higher than the threshold, and then the detection probability of the system can be calculated as follows:

order the

Then there are:

false alarm probability P of system _fa The threshold signal-to-noise ratio can be used to represent:

should keep T _h With the mean value m of background clutter _c And the noise root mean square sigma dynamic change can keep the constant false alarm rate of the system, so that a threshold value calculation formula can be obtained:

T _h ＝m _c +TNR·σ (17)

in order to facilitate design and adjustment, TNR is replaced by an inputtable parameter k in the subsequent design, different false alarm probabilities are set according to specific scenes, so that different k values are calculated for input, and a final threshold calculation formula is as follows:

T _h ＝m _c +k·σ (18)

and (3) carrying out threshold segmentation on the saliency map by using the constant false alarm rate, if the contrast ratio in the segmentation result is larger than the threshold value, proving that the region has high local contrast ratio, and outputting the region coordinates to finish infrared weak and small target detection.

Aiming at the two-layer efficient filtering templates, the parameters of the 2 layers of the saliency map are processed, and the pixels of the saliency map obtained by the two-layer efficient filtering templates are more and denser, so that the result is more accurate.

The infrared dim target detection method utilizes the characteristic that the infrared dim target has larger difference from the background and has higher contrast locally to detect the infrared dim target, and in the implementation process, the transmission bandwidth of a digital system is fully utilized through the high-efficiency filtering template, wherein the high-efficiency filtering template is a square template with the size of 2 and the number of pixels of even power, and the divided sub-blocks in the template are squares with the size of 2 and the number of pixels of power, so that the image traversing time can be reduced by using the template, and the system processing speed is greatly improved; meanwhile, the algorithm complexity is reduced, floating point operation and nonlinear operation are reduced, the processing time delay of the system is reduced, the design of a large-scale assembly line is facilitated, and the hardware implementation difficulty is reduced. Finally, the method can realize reliable high-speed infrared weak and small target detection, the throughput of the actually measured system reaches 20Gbps, the detection rate is higher than 90%, the false alarm rate is lower than 10%, the high-speed detection can be realized in engineering application, and the processing speed index is far higher than that of the similar algorithm.

Examples:

the invention is realized on a high-speed digital image processing board designed by an FPGA and DSP dual-processor architecture, the FPGA is a main processor model Kintex7, and is responsible for realizing most algorithm functions and peripheral chip control, the DSP chip is a coprocessor, and is responsible for managing processed data and communicating with an upper computer. The actual hardware platform and chip model are shown in fig. 5: the total processing rate is 20Gbps, the input of a 8X SRIO port of the system is equivalent to 8 paths of parallel processing, and each path requires the minimum rate of 2.5Gbps; the single-path SRIO adopts a working mode of 3.125Gbps, and the working main frequency is 125MHz; the caching chip adopts a DDR3 chip, the size is 2GB, the working speed is 1333MT/s, 20 pictures can be cached, the reading and writing operation clock frequency is 166MHz, and the reading and writing bit width is 512 bits.

In fig. 7, the specific implementation steps are as follows:

step 1: high-speed transmission of digital images;

image data acquired by the infrared camera is input into the system through a high-speed interface. The invention adopts 2 SRIO interfaces with 4X working modes, the single-port rate is set to be 3.125Gbps, and the highest throughput rate which can be achieved is 25Gbps. In order to meet the stability of high-speed data transmission and reduce transmission failure and rate loss caused by data retransmission, an SRIO interface generally needs a low-jitter high-precision clock, and the Si5368 clock chip is adopted to meet the requirement in the invention, and the Si5368 clock chip stably outputs a 125MHz high-precision clock.

Step 2: a cache of digital images;

the original infrared image is cached in a cache chip to wait for processing. The invention uses DDR3 chip to buffer data, 8 pieces DDR3 are altogether, the chip model is MT41J256M8HX_15E, and the control module is designed based on MIG IP core in VIVADO 2017.4.

Step 3: and (5) rapidly detecting the infrared weak and small target.

The infrared weak target step is as described in the detailed description.

Fig. 8, fig. 9 and fig. 10 show the processing results of the method in the practical application, fig. 8 represents the input original image, wherein 6 infrared dim targets in different backgrounds exist, fig. 9 is a significant graph output after contrast detection, and fig. 10 is a detection result output after threshold segmentation.

Further elaborating on efficient filtering templates

Distinction of efficient filter templates from other filter templates: the high-efficiency filtering templates are all 2-power-of-power pixels in length and width and are divided into 2-power sub-blocks. For example, the template in this application is 32×32 pixels in size, and then one sub-block is formed every 8×8 pixels. The common filtering templates generally do not satisfy the power of 2 pixels in length and width. A common filtering template is typically centrosymmetric and is typically divided into odd sub-blocks, e.g. 5 x 5, 7 x7, which results in that the length and width of the template must not be a power of 2 pixels. The first layer of the efficient filtering template is shown in fig. 2, and the second layer of the efficient filtering template is shown in fig. 3. In practical application, the efficient filtering template is 2 ⁿ ×2 ⁿ The square template of each sub-block is more than or equal to 2 and less than or equal to 5, the size of each sub-block is 8 multiplied by 8 pixels, the effective filtering template is necessarily square, the lower limit of 4 multiplied by 4 sub-blocks is at least capable of containing a processing unit for the efficient filtering template, and the upper limit of 32 multiplied by 32 sub-blocks is larger in resource consumption after the size of the template is increased, so that the effective filtering template is not practical any more. The specific size is configured with the actual situation of the digital system.

Fig. 14 shows a filtering template actually used in the present application, where a black square represents a pixel, a painted portion represents a sub-block, and three sub-blocks are divided, and a matrix of 32×32 pixels is transmitted in a digital system, which is a basic "transmission unit", and 3×3 sub-blocks in the template are taken out to form a "processing unit", where the processing unit is used to detect contrast, and a "transmission unit" includes 5 "processing units", where the actual correspondence is as shown in fig. 15:

the processing unit refers to a square area of 3×3 sub-block size (fixed value), as shown in fig. 16:

wherein area is ₀ ～area ₈ Representing 8 sub-blocks, the weak and small target occupying less than 9×9 pixels in the infrared image according to the international optical engineering council definition, so each sub-block is defined as 8×8 pixels in size, e.g. area ₀ The actual size of (a) is 8 x 8 pixels (fixed value). From this, it can be deduced that a processing unit must be 24 x 24 pixels in size.

The left side of fig. 15 represents transmission of a high-efficiency filtering template matrix in the system, the middle of the dotted line of fig. 15 represents a contrast measurement process, 3×3 sub-blocks are taken out from the high-efficiency filtering template to form a "processing unit", the contrast of the central gray area of each processing unit is calculated according to a contrast measurement algorithm, and the contrast of 5 small sub-blocks is obtained after calculation, such as the gray sub-blocks on the right side of fig. 15. The center of the processing unit is defined as area ₀ The background is defined as area ₁ ～area ₈ . The white background of fig. 15 represents the efficient filtering template, i.e. the transmission unit, and the color part represents one processing unit.

Saliency map generation process

Fig. 17 is a view showing the removal of one processing unit from the "transmission unit" to calculate the surrounding sub-block mean and the maximum/minimum value of the sub-block at the center. The resulting data may continue to calculate contrast for a certain direction and generate a saliency map, e.g. calculate contrast for sub-block 10:

upper direction contrast= (1/2) × (maximum value+next largest value) - (1/3) × (upper direction field pixel mean sum)

＝(1/2)×(200+195)-(1/3)×(36+96+66)

＝131.5

Lower contrast= (1/2) × (200+195) - (1/3) × (55+38+78)

＝140.5

First quadrant direction contrast= (1/2) × (200+195) - (1/3) × (96+66+48)

＝127.5

Similarly, the contrast in 8 directions is calculated:

left direction contrast=158.8, right direction contrast=133.5, second quadrant direction contrast=145.2, third quadrant contrast=158.2, fourth quadrant contrast=142.8.

Such a calculation process is present 5 times in total in one transmission unit.

The "transmission unit" is a sliding window in the original view from the scale of the whole figure, as shown in fig. 18:

the sliding window slides on the original image to obtain transmission units, and the 1 transmission unit divides 5 processing units for contrast detection and can obtain the contrast of 5 central areas in all directions.

A saliency map is then calculated. The saliency map is also a "map" consisting of pixels, introducing the calculation rules for each pixel:

after calculating the contrast in each direction of the sub-block 10 in fig. 5, the contrast in each direction may be taken as an absolute value and summed to obtain 1 pixel of the saliency map.

For example, the 10 th sub-block of the current sliding window corresponds to the pixel value of the saliency map, the upward direction contrast +|the downward direction contrast |+ …

＝|131.5|+|140.5|+|158.8|+|133.5|+|127.5|+|145.2|+|158.2|+|142.8|

＝1138

The absolute value must be taken because the calculated contrast in a certain direction is not necessarily positive.

Similarly,

sub-blocks

5, 6, 9 and 20 of the sliding window can calculate a pixel of the saliency map, and the saliency map is double-layered because the sliding window is double-layered. The 1 transmission unit can calculate 5 pixels of the saliency map. As shown in fig. 7.

The sliding window slides the distance of two sub-blocks (2×8=16 pixels) on the original image each time, so that 4 blue sub-blocks of the first layer can be seamlessly spliced into the first layer of the saliency map, and the second layer of the saliency map is gapped. As shown in fig. 20 and 21.

The white part in fig. 20 and 21 is the original image, the gray part is the calculated saliency map, and 1 pixel in the saliency map represents 8×8 pixels (1 sub-block) in the original image; FIG. 20 is a first layer of saliency maps, with the left dashed box representing 4 saliency map pixels calculated at position 1 of the slide window (gray patches representing 1 pixel of the saliency map), and the right dashed box representing 4 saliency map pixels calculated at position 2 after sliding the slide window by a distance of 2 sub-blocks; fig. 21 is the second layer of saliency maps, the 5 th saliency map pixel calculated for each sliding window. The correspondence is as shown in fig. 22:

after the saliency map is calculated, each pixel of the saliency map is compared with a threshold value.

It should be noted that the detailed description is merely for explaining and describing the technical solution of the present invention, and the scope of protection of the claims should not be limited thereto. All changes which come within the meaning and range of equivalency of the claims and the specification are to be embraced within their scope.

Claims

1. A low-complexity infrared dim target detection method based on local contrast is characterized by comprising the following steps of: the method comprises the following steps:

the contrast detection comprises the following steps:

2. The method for detecting the low-complexity infrared dim target based on the local contrast according to claim 1, wherein the efficient filtering template is a double-layer template, the efficient filtering template comprises a first-layer template and a second-layer template, and the first-layer template is 2 ⁿ ×2 ⁿ Square templates of sub-blocks, the second layer template is 2 ⁿ -1×2 ⁿ -a square template of 1 sub-block, every four adjacent sub-blocks in the first layer template corresponding to one sub-block in the second layer template, the intersection points of the four adjacent sub-blocks in the first layer template coinciding with the center points of the sub-blocks in the second layer template.

3. The method for detecting the low-complexity infrared dim target based on the local contrast according to claim 2, wherein the specific steps of the contrast detection are as follows:

4. A low-complexity infrared small target detection method based on local contrast according to claim 3, characterized in that the saliency luminance Lmax of the target area is expressed as:

5. The method for detecting a low-complexity infrared dim target based on local contrast according to claim 4, wherein said background region has a significant luminance Lmean _i Expressed as:

6. The method for detecting low-complexity infrared dim target based on local contrast according to claim 5, wherein said salient luminance Lneighbor in each direction is characterized by _m Expressed as:

7. a method of detecting a low-complexity infrared small target based on local contrast according to claim 6, wherein the contrast measure is expressed as:

contrast _m ＝Lmax-Lneighbor _m

where m=1, 2, …,8.

8. A low-complexity infrared small target detection method based on local contrast according to claim 1, characterized in that the threshold value is expressed as:

T _h ＝m _c +k·σ

9. The method for detecting the low-complexity infrared dim target based on the local contrast according to claim 1, wherein the infrared image is acquired by an infrared camera.