A background adaptive suppression method for infrared dim small target detection

Abstract

The present invention discloses a background adaptive suppression method for infrared weak and small target detection. The method comprises the following steps: firstly performing a multi-scale dilation operation on an input single-frame infrared image, then performing median filtering on the enhanced results to suppress discrete noise points, then splicing multiple filtering results according to the image channel direction, selecting the maximum value according to the image channel direction as the pixel value of the current coordinate, obtaining a preprocessed one-channel image, performing block processing on the image, and calculating the fluctuation rate of the image and each image block. If the fluctuation rate of the current image block is less than the fluctuation rate of the entire image, it is highly likely that the image block is a background area. If the fluctuation rate of the current image block is greater than twice the fluctuation rate of the entire image, it is highly likely that the target area exists. The method of the present invention suppresses the complex background, highlights the real target, and realizes accurate detection of infrared weak and small targets.

Description

Technical Field

[0001] The present invention belongs to the technical field of image target detection, and in particular relates to a background adaptive suppression method for infrared dim small target detection. Background Art

[0002] Infrared sensor imaging is highly sensitive to a target's thermal radiation characteristics and is insensitive to changes in scene brightness. Therefore, it can effectively overcome the effects of poor lighting, rain, snow, wind, and frost, and has a wide range of applications. However, when a target is camouflaged or its thermal radiation characteristics are unclear, the number of effective pixels is low, resulting in a low signal-to-noise ratio. The true target in the image can easily be obscured by the background, leading to detection failure.

[0003] At present, many methods for background suppression in infrared images have been proposed, such as the top-hat transform algorithm based on morphological filtering, median and mean filtering methods, wavelet transform method, frequency domain high-pass filtering method, bilateral filtering method, etc.

[0004] These methods are all based on full-image processing of the grayscale or distribution of other features of the image. Therefore, when the background is complex or the target brightness is lower than the surrounding interference (such as buildings, leaves, and water reflections), they often cannot achieve a good target highlighting effect. Summary of the Invention

[0005] In response to the technical problems existing in the prior art, this paper proposes a background adaptive suppression method for infrared dim small target detection in complex environments. By suppressing the complex background, the real target is highlighted, and accurate detection of infrared dim small targets is achieved.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a background adaptive suppression method for infrared weak target detection, comprising the following steps

[0007] S1, preprocessing the input single-frame infrared image: setting the expansion operator of three scale structures, performing multi-scale image expansion operation on the input image, and performing expansion operation according to the following expansion formula

[0008]

[0009] Where I is the image of one channel of the original single-frame infrared image, K n is a structural element, is the expansion operator symbol, Φ represents the empty set;

[0010] S2, the expanded image is recorded as I_dilate, and the enhancement operator H is set to Perform sliding window convolution operation on the enhancement operator H and the image I_dilate, with a step size of 1 pixel each time. The image processed by the operator sliding window is superimposed on the image I_dilate to highlight the contour edge of the target, and the enhanced image I_dilate_enhance is obtained;

[0011] S3, the image I_dilate_enhance processed in the previous step is once again subjected to median filtering using a median filter to suppress some discrete noise points, thereby obtaining the image I_dilate_enhance_median;

[0012] S4, the image I_dilate_enhance_median processed in the previous step is stitched in the direction of the image channel to obtain a three-channel image with the same resolution as the original input image, and the maximum value of the channel in the three-channel image is selected, and the obtained one-channel image is recorded as I_dilate_enhance_median_cat;

[0013] S5, dividing the image I_dilate_enhance_median_cat into blocks, and dividing the blocks into blocks according to the following loop:

[0014] for j=1:width

[0015] for i=1:height

[0016] patch s =(1+i,1+j),

[0017] Where width is the width of the image, height is the height of the image, patch is the block obtained each time, subscript s is the total number of blocks that can be obtained in the end, and the boundary of the image is filled with the surrounding neighborhood pixel values;

[0018] S6, calculate the volatility of the image I_dilate_enhance_median_cat. At the same time, the volatility of each block obtained in the previous step is calculated separately. Here, the volatility is defined as:

[0019]

[0020] Among them, image max is the maximum grayscale value of the image pixel, image min is the minimum grayscale value of the image pixel, image i,j is the gray value at the coordinate (i, j), mean is the average value;

[0021] S7, respectively determine the fluctuation rate of each image block and the image I_dilate_enhance_median_cat: if the fluctuation rate of the current image block is less than the fluctuation rate of the entire image, it is considered that the grayscale distribution of the image in the current image block is relatively uniform and consistent, and it is very likely to be the background area; if the fluctuation rate of the current image block is greater than twice the fluctuation rate of the entire image, it is considered that the grayscale fluctuation in the current image block is large, and it is very likely to be the target area;

[0022] The criteria for determining whether the current image block is the background are:

[0023]

[0024] If it is a background area, the mean of the image block is calculated, and all pixel grayscale values ​​are replaced by the calculated mean;

[0025] If it is the target area, no processing is performed and the infrared weak target is obtained after background suppression.

[0026] Furthermore, the method further comprises the steps of:

[0027] S8, stitching each processed image block according to the block division rule in step S5 in the following cycle:

[0028] for j=1:width

[0029] for i=1:height

[0030] patch s =(1+i,1+j),

[0031] After restoration, a stitched image with the same resolution and one-to-one correspondence as the original input image is obtained. The average of the overlapping parts is used as the grayscale value of the restored position. The restored image is recorded as I_dilate_enhance_median_cat_adapfilter;

[0032] S9, then, calculate the mean m and standard deviation std of the image I_dilate_enhance_median_cat_adapfilter, set the binarization threshold T = m + 5*std, mark the pixel positions greater than the threshold as 1, and the rest as 0, to obtain a binarized image, and then calculate the centroid of the binarized image as the target location.

[0033] Furthermore, the dilation operator in step S1 is a 5×5, 9×9 and 17×17 scale structure.

[0034] Furthermore, in step S3, the median filter processing window is 9*9.

[0035] Furthermore, in the block processing in step S5, the moving step length for each time is 5.

[0036] The beneficial effects of the present invention are as follows: the method of the present invention first pre-processes the image, then divides the pre-processed image into blocks, performs adaptive local filtering on each block to suppress the background, re-joins each processed block according to certain rules, sets an adaptive threshold to perform image binarization, and realizes target extraction; the method of the present invention follows the idea of ​​block division, performs divide-and-conquer filtering processing on different image blocks according to image fluctuation rates, and adaptively suppresses the background, effectively solving the difficult problem of difficult detection of infrared weak targets under complex backgrounds. DETAILED DESCRIPTION

[0037] The present invention will be further described in detail below with reference to the embodiments.

[0038] The present invention discloses an infrared small target detection background suppression method based on adaptive local filtering, which includes the following steps.

[0039] S1, preprocessing the original input single frame infrared image.

[0040] Since the three RGB channels of an infrared image contain identical information, any one of the three channels is arbitrarily selected for processing. Since the specific size of the target is uncertain, a multi-scale image morphological dilation operation is required. This dilation operation expands boundary points, merging all background points that touch the object into the object, expanding the boundary outward. This is useful for stretching the size of small targets and, to a certain extent, connecting discrete noise points to facilitate further processing.

[0041] Generally speaking, the size of a typical small object in an image ranges from 2×2 to 11×11 pixels. Based on this, this patent sets three structural elements, and the 17×17 operation kernel is as follows:

[0042] 0 0 0 0 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 1 1 1 1 1 1 1 1 1 1 1 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1 0 0 0 0 .

[0043] The 9×9 operation core is as follows:

[0044] 0 0 1 1 1 1 1 0 0 0 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 0 1 1 1 1 1 1 1 0 0 0 1 1 1 1 1 0 0 .

[0045] The 5×5 operation kernel is as follows:

[0046] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 .

[0047] According to the following image processing expansion formula, the expansion operation is performed separately:

[0048]

[0049] Where I is the image of one channel of the original single-frame infrared image, K n is a structural element, is the dilation operator symbol, Φ represents the empty set; a dilation result is obtained for each scale, and then the enhancement operator is used to perform enhancement operations on the multi-scale dilation results respectively. Since the dilation operation is a basic concept in digital image processing, it will not be repeated here.

[0050] S2, the expanded image is recorded as I_dilate. At this time, the size of the target has been magnified at multiple scales, but isolated noise points or other background objects will also be magnified.

[0051] Infrared small targets generally show a Gaussian distribution characteristic with high center and low surrounding in the image, which is a maximum value in its field. Therefore, in this step, the enhancement operator H needs to be set as The enhancement operator H is convolved with the image I_dilate using a sliding window convolution operation. The step size of each sliding window is 1 pixel. The image processed by the operator sliding window is superimposed on the image I_dilate to highlight the contour edge of the target, and the enhanced image I_dilate_enhance is obtained.

[0052] S3, the image I_dilate_enhance processed in the previous step is once again subjected to median filtering to suppress some discrete noise points, where the processing window of the median filter is 9*9, and the processed image is recorded as I_dilate_enhance_median.

[0053] S4. Through the preprocessing of steps S1 to S3 above, the discrete and isolated noise points in the background can be removed, and a total of three different results are obtained. These three results are spliced ​​in the direction of the image channel to obtain a three-channel image with the same resolution as the original input. For each pixel in the image, the maximum value of the three channels of the three-channel image is selected. The final one-channel image is recorded as I_dilate_enhance_median_cat.

[0054] S5, then, the image I_dilate_enhance_median_cat is divided into blocks. In order to balance the computation time and the target size, and to avoid the gradient changes at the block boundaries after subsequent processing, the blocks are divided according to the following loop:

[0055] for j=1:width

[0056] for i=1:height

[0057] patch s =(1+i,1+j),

[0058] Where width is the width of the image I_dilate_enhance_median_cat, height is the height of the image I_dilate_enhance_median_cat, patch is the block obtained each time, the subscript s is the total number of blocks that can be obtained in the end, the moving step size of each time is 5, and the image boundary is filled with the surrounding neighborhood pixel values.

[0059] S6, calculate the volatility of the image I_dilate_enhance_median_cat. At the same time, the volatility of each block obtained in the previous step is calculated separately. Here, the volatility is defined as:

[0060]

[0061] Among them, image max is the maximum grayscale value of the image pixel, image min is the minimum grayscale value of the image pixel, image i,j is the grayscale value at the coordinate (i, j), mean is the average value, and the volatility reflects the degree of fluctuation of the current image.

[0062] S7 , respectively, determines the fluctuation ratio of each image block to the image I_dilate_enhance_median_cat. If the fluctuation ratio of the current image block is less than the fluctuation ratio of I_dilate_enhance_median_cat for the entire image, the grayscale distribution of the current image block is considered relatively uniform and consistent, and it is likely to be a background area. If the fluctuation ratio of the current image block is greater than twice the fluctuation ratio of the entire image, the grayscale fluctuations in the current image block are considered to be large, and it is likely to contain a target area. Here, the fluctuation ratio is positively correlated with the image background; a larger fluctuation ratio indicates a higher probability of noise.

[0063] The criteria for determining whether the current image block is the background are:

[0064]

[0065] If it is a background area, the mean of the image block is calculated, and all pixel grayscale values ​​are replaced by the calculated mean; if it is a target area, no processing is performed, and an infrared weak target after background suppression is obtained.

[0066] S8, the operation in the previous step, according to the adaptive local processing method, can largely remove background interference. Then, each image block after processing is spliced ​​according to the block division rule in step S5, according to the following cycle:

[0067] for j=1:width

[0068] for i=1:height

[0069] patch s =(1+i,1+j),

[0070] Since some areas of adjacent image blocks in step 5 are overlapping, the mean of the overlapping parts is calculated as the grayscale value of the restored position when restoring. The restored image is recorded as I_dilate_enhance_median_cat_adapfilter.

[0071] S9, then calculate the mean m and standard deviation std of the image I_dilate_enhance_median_cat_adapfilter, set the binarization threshold T = m + 5 * std, mark the pixel positions greater than the threshold as 1, and the rest as 0, to obtain a binarized image, and then calculate the centroid of the binarized image as the target position, and finally obtain the infrared weak target after background suppression.

[0072] The above embodiments are merely illustrative of the principles and effects of the present invention, as well as some embodiments of its application. A person skilled in the art may make several modifications and improvements without departing from the inventive concept of the present invention, and all of these modifications and improvements fall within the scope of protection of the present invention.

Claims

1. A background adaptive suppression method for infrared small target detection, characterized by: Includes the following steps S1, preprocessing the input single-frame infrared image: setting the expansion operator of three scale structures, performing multi-scale image expansion operation on the input image, and performing expansion operation according to the following expansion formula Where I is the image of one channel of the original single-frame infrared image, K n is a structural element, is the expansion operator symbol, Φ represents the empty set; S2, the expanded image is recorded as I_dilate, and the enhancement operator H is set to Perform sliding window convolution operation on the enhancement operator H and the image I_dilate, with a step size of 1 pixel each time. The image processed by the operator sliding window is superimposed on the image I_dilate to highlight the contour edge of the target, and the enhanced image I_dilate_enhance is obtained; S3, using a median filter to perform median filtering on the image I_dilate_enhance to suppress some discrete noise points, thereby obtaining an image I_dilate_enhance_median; S4, concatenate the image I_dilate_enhance_median in the direction of the image channel to obtain a three-channel image with the same resolution as the original input image, select the maximum channel value in the three-channel image, and the obtained one-channel image is recorded as I_dilate_enhance_median_cat; S5, dividing the image I_dilate_enhance_median_cat into blocks, and dividing the blocks into blocks according to the following loop: for j=1:width for i=1:height patch s =(1+i,1+j), Where width is the width of the image, height is the height of the image, patch is the block obtained each time, subscript s is the total number of blocks that can be obtained in the end, and the boundary of the image is filled with the surrounding neighborhood pixel values; S6, calculate the image I_dilate_enhance_median_cat and the volatility of each image block using the following formula: Among them, image max is the maximum grayscale value of the image pixel, image min is the minimum grayscale value of the image pixel, image i,j is the gray value at coordinate (i, j), mean is the average value; S7, respectively judge the volatility of each image block and the image I_dilate_enhance_median_cat: if the volatility of the current image block is less than the volatility of the entire image, it is very likely that it is a background area; if the volatility of the current image block is greater than twice the volatility of the entire image, it is very likely that there is a target area; if it is a background area, calculate the mean of the image block, and replace all pixel grayscale values ​​with the calculated mean; if it is a target area, obtain an infrared weak target after background suppression.

2. The background adaptive suppression method for infrared dim small target detection according to claim 1 is characterized in that: Also includes the steps: S8, stitch each processed image block according to the following loop: for j=1:width for i=1:height patch s =(1+i,1+j), Get a stitched image with the same resolution and one-to-one correspondence as the original input image. The average of the overlapping parts is used as the grayscale value of the restored position. The restored image is recorded as I_dilate_enhance_median_cat_adapfilter; S9, calculate the mean m and standard deviation std of the image I_dilate_enhance_median_cat_adapfilter, set the binarization threshold T = m + 5*std, mark the pixel positions greater than the threshold as 1, and mark the rest as 0, to obtain a binarized image, and then calculate the center of mass of the binarized image as the target location.

3. The background adaptive suppression method for infrared dim small target detection according to claim 1 or 2, characterized in that: The dilation operators in step S1 are three scale structures: 5×5, 9×9, and 17×17.

4. The background adaptive suppression method for infrared dim small target detection according to claim 1 or 2, characterized in that: The median filter processing window in step S3 is 9*9.

5. The background adaptive suppression method for infrared dim small target detection according to claim 1 or 2, characterized in that: In the block processing in step S5, the moving step length for each time is 5.

Images