A small target detection method based on sea-sky line

Abstract

The application discloses a small target detection method based on a sea-sky line, and comprises the following steps: firstly, a region complexity model based on single-pixel complexity, region gray scale and image signal-to-noise ratio is established; then, the image is segmented according to scene complexity; then, region complexity statistics and spatial suppression are sequentially performed; according to the region complexity and the spatial suppression result, the image is divided into regions according to a set threshold; after a binary image is obtained, feature extraction is performed; finally, small targets based on Tophat are detected. The image is segmented according to scene complexity, and no target is extracted in a complex region, so that the purpose of eliminating false alarms is achieved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and infrared image processing, and particularly relates to a small target detection method based on a sea-sky line. BACKGROUND

[0002] In the field of image processing and computer vision, the generalization of algorithms is constrained by scenes. In different scenes, various image algorithms often fail to achieve good results, such as a large number of false alarm targets generated in the detection task. These false alarm targets not only increase the difficulty of subsequent processing, but also may interfere with the final target recognition result. Therefore, scene recognition has become one of the important preprocessing links for subsequent image tasks.

[0003] In order to solve the above problems, a large number of researches have been conducted in recent years. According to the output results, the methods can be divided into two categories. The first category is a method of directly outputting scene categories, such as a scene classification method based on traditional image signals, image features and support vector machines, an image classification model PLACECNN based on deep learning, etc. The second category is a segmentation method of dividing the input scene into regional categories, such as a K-means clustering method based on clustering methods, a semantic segmentation method based on deep learning, etc. Both of the two methods have their shortcomings. The method of directly outputting image types enables the entire algorithm framework to add more corresponding strategies to adapt to the actual scene, but it is difficult to add effective hard constraints to the output results of target detection and target segmentation methods. The method of outputting scene segmentation regions can more detailedly express the categories to which the pixels belong, and facilitate the filtering of subsequent vision tasks such as detection tasks, but it brings greater computational load and requires stronger and more robust feature expression.

[0004] Whether the scene segmentation method is based on support vector machines or deep learning, data needs to be trained, which is called a data-driven method. This kind of method will bring huge computational load, and it is difficult to deploy on edge devices such as embedded devices and FPGA. The method based on traditional image feature statistics relies on prior knowledge, and needs more strategies to adapt to the scene. SUMMARY

[0005] In view of the high false alarm rate of small target detection and segmentation on the sea surface, the application proposes a small target detection method based on a sea-sky line, which assists small target detection with scene segmentation of the sea-sky background.

[0006] In order to achieve the above purpose, the technical scheme adopted by the application to solve the technical problems is: a small target detection method based on a sea-sky line, comprising the following steps:

[0007] S1, region complexity modeling: a region complexity model RC=F(C, RG, SNR) is established based on single-pixel complexity C, region gray RG and image signal-to-noise ratio SNR, wherein the single-pixel complexity C is used to measure the difference between the pixel and the neighborhood, the region gray RG is used to reflect the brightness characteristics of the region, and the image signal-to-noise ratio SNR is used to reflect the proportional relationship between noise and signal in the image;

[0008] S2, a horizon line posture coordinate is introduced in the image, the image is segmented into a complex region and a flat region according to scene complexity, the complex region includes a complex cloud layer region and a complex sea surface region, and the flat region includes a flat sky background region and a flat sea surface region, so as to realize fine segmentation of the image scene;

[0009] S3, the image is traversed using an overlapping sliding window strategy, and the region complexity in each window is calculated; the brightness influence of the target on the image is eliminated by an opening operation of a large window, only the gray mean value of the background is retained, then the morphological filtering is performed on the down-sampled image according to the image pyramid idea, the background mean value is obtained and normalized, the horizon line posture coordinate is taken as the demarcation line, the standard deviation of the sea surface and the sky is calculated respectively as the stretching strength to complete the spatial suppression; according to the region complexity and the spatial suppression result, the image is regionally divided according to the set threshold value, and a binary image is obtained;

[0010] S4, the divided binary image is subjected to region connection and isolated point elimination processing to obtain the final scene segmentation result;

[0011] S5, small target detection based on Tophat

[0012] S51, the horizon line of the input image is calibrated, and then TopHat transformation preprocessing and image scaling are respectively performed;

[0013] S52, the image after TopHat preprocessing in step S51 is subjected to threshold segmentation and overlapping window region extraction;

[0014] S53, the image after overlapping window region extraction in step S52 is subjected to region statistics;

[0015] S54, the scaled image in step S51 is subjected to opening operation and normalization in sequence, then is subjected to AND operation (spatial point multiplication) with the picture after region statistics, then is subjected to threshold segmentation and morphological filtering in sequence, and then is subjected to region suppression with the picture after threshold segmentation in step S52;

[0016] S55, the minimum bounding box is found, and the target information is output.

[0017] Further, the complex region and the flat region are taken as the highlight region and the low-light region respectively, and according to the significant difference between the highlight region and the low-light region in the gray scale, the image region brightness after down-sampling is taken as the space suppression term, and the highlight region is given higher complexity.

[0018] Further, the image signal-to-noise ratio SNR is processed by prior knowledge to process the influence of the signal-to-noise ratio.

[0019] Further, in the step S3, the mean value of multiple calculations is used to complete the region complexity statistics for the overlapping region.

[0020] Further, the region suppression part in the step S54 can be operated in two types, one is region zeroing to eliminate the target of the complex region, and the other is to suppress the target weight to suppress the less important target, so as to achieve the re-screening of the target detection.

[0021] The present application has the advantages that: the present application introduces the sea-sky line posture coordinates in the image, divides the image into the complex region and the flat region according to the scene complexity, realizes the fine segmentation of the image scene, adopts different processing strategies according to the characteristics of each sub-region to reduce the false alarm target in the target detection process, divides the image according to the scene complexity to achieve the purpose of false alarm elimination, assists the small target detection through the scene segmentation, and applies the small target detection method based on Tophat to realize feature reuse and effectively reduce the calculation amount. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 It is a scene segmentation framework of the present application;

[0023] Figure 2 It is a comparison diagram of two windowing methods of the present application;

[0024] Figure 3 It is a flowchart of the small target detection method based on Tophat of the present application. DETAILED DESCRIPTION

[0025] The present application is further described in combination with the drawings and embodiments as follows.

[0026] The method of the present application divides the image according to the scene complexity, and does not extract the target in the complex region to achieve the purpose of false alarm elimination. Therefore, first, the definition and characteristics of various scenes in the image need to be clarified, such as the high brightness of the complex cloud layer and the texture complexity of the complex sea surface. At the same time, according to the task requirement, the image segmentation task is defined as a semantic segmentation task, and the scene segmentation process of region division-feature extraction-post processing is determined, as shown in Figure 1 .

[0027] The small target detection method based on the sea-sky line disclosed in the present application comprises the following steps.

[0028] S1, establishing a region complexity model.

[0029] In order to quantify the region complexity of the image, the present application introduces three factors of single pixel complexity C, region gray RG and image signal-to-noise ratio SNR, and establishes a region complexity model RC=F(C,RG,SNR). The single pixel complexity C is used to measure the difference between the pixel and the neighborhood; the region gray RG reflects the brightness characteristics of the region; and the image signal-to-noise ratio SNR reflects the proportional relationship between the noise and the signal in the image.

[0030] For the image signal-to-noise ratio SNR, since it is difficult to directly obtain the image signal-to-noise ratio, the present application processes the influence of the signal-to-noise ratio through prior knowledge. The effect of the signal-to-noise ratio mainly reflects in the threshold setting of scene segmentation and target extraction. Therefore, the present application interfaces the signal-to-noise ratio for adjusting the value.

[0031] S2, segmenting the image according to the scene complexity.

[0032] The horizon posture coordinates are introduced in the image, the image is segmented into complex regions and flat regions according to the scene complexity, and is further subdivided into complex cloud layer regions, complex sea surface regions, flat sky background regions and flat sea surface regions, so as to realize the fine segmentation of the image scene. Meanwhile, according to the characteristics of each sub-region, different processing strategies are adopted to reduce the false alarm targets in the target detection process.

[0033] For the region gray RG, considering the significant difference in gray between the high-brightness regions (such as the complex cloud layer region and the sea surface bright band region) and the low-brightness regions (such as the flat sky background region and the flat sea surface region), the present application uses the brightness of the down-sampled image region as a spatial suppression term to give higher complexity to the high-brightness regions.

[0034] For the single pixel complexity C, the present application uses the result of the Tophat algorithm to measure the difference between the pixel and the neighborhood. The Tophat algorithm can highlight the high-brightness regions and details in the image, so that the scene segmentation can be coupled with the original target extraction to reduce repeated calculation.

[0035] According to the region complexity model, the scene segmentation is performed by using the method of combining region complexity statistics and spatial suppression. The specific steps are as follows.

[0036] S3, region complexity statistics: using the overlapping sliding window strategy to traverse the image, calculating the region complexity in each window, and using the mean value of multiple calculations to complete the region complexity statistics for the overlapping regions. The method of overlapping sliding window is as follows Figure 2As shown, the hard division can alleviate the problem of uneven segmentation results caused by hard division, and reduce the hole phenomenon of the segmented region. For the overlapping region, the mean value of multiple calculations is used to obtain the final region complexity. Figure 2 As shown, the left side is the hard windowing mode, and the right side is the overlapping windowing mode.

[0037] Spatial inhibition: Since there are deficiencies in using only complexity features for region division, such as the high-light region and the low-noise flat region may be misclassified as the same situation, the present application introduces a spatial inhibition branch to solve this problem. The spatial inhibition branch eliminates the brightness influence of the target on the image by the opening operation of the large window, only retains the gray mean value of the background, and then uses the image pyramid idea to perform morphological filtering on the down-sampled image to obtain the background mean value and perform normalization processing. In the normalization process, the present application takes the sea-sky line as the demarcation line, respectively calculates the standard deviation of the sea surface and the sky as the stretching strength, so as to establish the normalization calculation process.

[0038] According to the region complexity and the spatial inhibition result, the image is regionally divided according to the set threshold value to obtain a binary image.

[0039] S4, feature extraction. The divided binary image needs to be processed by region connection and isolated point elimination to obtain the final scene segmentation result.

[0040] S5, post-processing.

[0041] Small target detection based on Tophat: Since the region complexity needs to be extracted, the present application adopts the Tophat small target detection method, which can well achieve feature reuse and effectively reduce the calculation amount.

[0042] The present application proposes a small target detection framework based on Tophat, as shown in Figure 3 As shown, the present application needs to extract the region complexity, and the Tophat small target detection method can well achieve feature reuse and effectively reduce the calculation amount. The specific steps are as follows.

[0043] S51, sea-sky line calibration is performed on the input image, and then TopHat transformation preprocessing and image scaling are respectively performed.

[0044] S52, the image after TopHat preprocessing in step S51 is respectively subjected to threshold segmentation and overlapping window region extraction.

[0045] S53, region statistics is performed on the image after overlapping window region extraction in step S52.

[0046] S54, the image scaled in step S51 is sequentially subjected to an open operation and normalization, and then is subjected to a spatial vector point multiplication with the picture subjected to region statistics, and then is sequentially subjected to threshold segmentation and morphological filtering, and then is subjected to region suppression with the picture subjected to threshold segmentation in step S52.

[0047] The region suppression part can perform two types of operations, one is region zeroing to eliminate complex regions of the target, and the other is to suppress the target weight to suppress the target with lower importance, so as to achieve re-screening of target detection.

[0048] S55, find the minimum bounding box, and output the target information.

[0049] The weight of the small target detection part, the weight of the scene segmentation part, and the calibration of the sea-sky line in the image coordinates can be manually intervened by the host computer, or a relative threshold can be set to automatically close the loop.

[0050] In order to better deploy on the edge device, the present application adopts a 7x9 size morphological operation to realize this process except for the morphological filtering of the scene segmentation part before region suppression, which has obvious advantages in operator optimization and implementation of the edge device. After threshold segmentation of the scene segmentation part, the morphological filtering needs to connect the segmentation point set and then delete the isolated noise points, so it needs to be morphologically dilated first and then combined with a larger size morphological erosion operator. Here, the erosion operator adopts 7x9, and the dilation operator adopts 5x5 size, which effectively connects the regions and removes isolated noise points.

[0051] The above embodiments only exemplarily illustrate the principles and effects of the present application, and part of the applied embodiments, and those skilled in the art can make several modifications and improvements without departing from the concept of the present application, which all belong to the protection scope of the present application.

Claims

1. A method for detecting small targets based on sea-sky line, characterized in that: The method comprises the following steps S1, establishing a region complexity model RC=F(C, RG, SNR) based on single-pixel complexity C, region gray RG and image signal-to-noise ratio SNR, wherein the single-pixel complexity C is used to measure the difference between a pixel and its neighborhood, the region gray RG is used to reflect the brightness characteristics of the region, and the image signal-to-noise ratio SNR is used to reflect the proportional relationship between noise and signal in the image; S2, introducing a horizon line posture coordinate in the image, and dividing the image into a complex region and a flat region according to scene complexity, wherein the complex region comprises a complex cloud layer region and a complex sea surface region, and the flat region comprises a flat sky background region and a flat sea surface region; S3, using an overlapping sliding window strategy to traverse the image, calculating the region complexity in each window, eliminating the brightness influence of the target on the image through an opening operation of a large window, retaining only the gray mean value of the background, then performing morphological filtering on the down-sampled image by using an image pyramid idea, obtaining the background mean value and performing normalization processing, taking the horizon line posture coordinate as a demarcation line, calculating the standard deviation of the sea surface and the sky as the stretching strength respectively to complete spatial suppression, and performing region division on the image according to the region complexity and the spatial suppression result to obtain a binary image; S4, performing region connection and isolated point elimination processing on the divided binary image to obtain a final scene segmentation result; S5, small target detection based on Tophat S51, calibrating the horizon line of the input image, then performing TopHat transformation and image scaling respectively; S52, performing threshold segmentation and overlapping window region extraction on the image preprocessed by TopHat respectively; S53, performing region statistics on the image after overlapping window region extraction; S54, sequentially performing an opening operation and normalization on the scaled image, then performing spatial vector point multiplication on the image with the region statistics picture, then sequentially performing threshold segmentation and morphological filtering, and then performing region suppression on the picture after threshold segmentation in step S52; S55, finding a minimum bounding box and outputting target information.

2. The small target detection method based on the horizon line according to claim 1, characterized in that, The complex region and the flat region are taken as a high-brightness region and a low-brightness region respectively, the difference between the high-brightness region and the low-brightness region in gray is adopted, the region brightness of the down-sampled image is taken as a spatial suppression term, and the high-brightness region is given a complexity.

3. The small target detection method based on the horizon line according to claim 1, characterized in that, The image signal-to-noise ratio SNR is processed by prior knowledge to process the influence of the signal-to-noise ratio.

4. The small target detection method based on the horizon line according to claim 1 or 2 or 3, characterized in that, In step S3, the mean value of multiple calculations is used to complete region complexity statistics for overlapping regions.

5. The small target detection method based on the horizon line according to claim 4, characterized in that, In step S54, the target in the complex region is eliminated by region zeroing for region suppression, or the target with low importance is suppressed by target weight suppression, so as to achieve re-selection of target detection.

Images