An infrared ship target detection method for complex sea-sky background

Abstract

The application relates to an infrared ship target detection method for a complex sea-sky background, which comprises the following steps: constructing an Adaboost ship detection classifier, training the Adaboost ship detection classifier based on HOG features, realizing rough ship detection, expanding a detection area, carrying out block row and column mean enhancement on the detection area, using step-by-step image segmentation on the enhanced image, and obtaining a target detection result through the segmentation result. The application uses an image processing method to detect the infrared ship target, has low calculation amount, is convenient for transplantation on an embedded platform, and can obtain high-precision detection results in various complex sea-sky backgrounds and has excellent robustness.

Description

Technical Field

[0001] This invention relates to the field of image processing target detection, specifically to an infrared ship target detection method for complex sea and sky backgrounds. Background Technology

[0002] Infrared target detection against a sea-sky background is of great significance in maritime transportation and other fields, and has received unprecedented attention. The accuracy and false alarm probability of target detection affect the efficiency of tasks such as ship positioning.

[0003] Visible light imaging severely impacts image quality when light sources are insufficient or targets are exposed to adverse weather conditions. Any object with a temperature above absolute zero emits thermal radiation. Infrared imaging technology uses photoelectric techniques to receive the infrared signals of this thermal radiation, thereby displaying a human-readable image in the imaging system. In terms of detection algorithms, target detection in complex scenes is a key research area, with numerous scholars both domestically and internationally conducting extensive work and achieving significant results. However, due to the low contrast of ships in infrared images, limited usable information, and interference from sea surface glare and other factors, infrared ship target detection in complex scenes remains a bottleneck in infrared target detection technology. It is a research hotspot in the fields of optics and infrared imaging, and a core technology for infrared search and tracking systems, large field-of-view infrared target surveillance systems, and more.

[0004] Low-contrast target detection in strong clutter backgrounds directly affects the detector's operating range and system detection performance. There are relatively few target detection algorithms in this area, and directly applying existing target detection algorithms to infrared ship detection does not yield good results. Summary of the Invention

[0005] This invention addresses the shortcomings of existing technologies by proposing an infrared ship target detection method for complex sea and sky backgrounds. This method effectively solves the problems of false detection and missed detection of infrared ship targets caused by complex backgrounds, low contrast, unclear target outlines, and few texture details, thereby improving the detection and recognition capabilities of infrared ship targets.

[0006] The technical solution adopted by this invention to achieve the above objectives is: an infrared ship target detection method for complex sea and sky backgrounds, comprising the following steps:

[0007] Step 1: Construct the Adaboost ship detection classifier; the input image is passed through the Adaboost ship detection classifier to generate coarse ship detection boxes;

[0008] Step 2: Expand the coarse detection frame area;

[0009] Step 3: Perform block-based row and column mean enhancement on the expanded coarse detection box region to obtain the enhanced image;

[0010] Step 4: Perform step-by-step image segmentation on the enhanced image;

[0011] Step 5: Obtain the target detection result through the segmentation result.

[0012] In step 1, the Adaboost ship detection classifier uses an adaptive weight adjustment method to train the base classifier. In each iteration, the weights of the training samples are adjusted according to the previous classification error rate to gradually enhance the weak classifier.

[0013] In step 2, the height of the coarse detection frame area is increased by m times and the width is increased by n times, resulting in an expanded area with height h0 and width w0.

[0014] In step 3, the expanded coarse detection box region is subjected to block-based row and column mean enhancement, which includes the following steps:

[0015] Gaussian filtering is applied to the expanded coarse detection box region image to obtain image I0. Image I0 is then divided into left and right blocks and top and bottom blocks respectively.

[0016] The left and right division includes the following steps:

[0017] (5.1) Take the leftmost region with height h0 and width 1 / k×w0, and calculate the mean of each row to obtain an h0×1 dimension vector; k is the coefficient;

[0018] (5.2) Take an empty matrix of size h0×w0, fill each column with the vector obtained in the previous step, and obtain the left block image;

[0019] (5.3) Take the rightmost region with height h0 and width 1 / k×w0, and calculate the mean of each row to obtain the h0×1 dimension vector;

[0020] (5.4) Take an empty matrix of size h0×w0, fill each column with the vector obtained in the previous step, and obtain the right block image;

[0021] (5.5) Subtract the left block image and the right block image from the image I0 respectively, and perform Gaussian filtering respectively to obtain the enhanced image of the left block image and the enhanced image of the right block image.

[0022] The upper and lower blocks include the following steps:

[0023] (6.1) Take the region with a top height of 1 / k×h0 and a width of w0, and calculate the mean of each column to obtain a 1×w0 dimensional vector; k is the coefficient;

[0024] (6.2) Take an empty matrix of size h0×w0, and fill each row with the vector obtained in the previous step; to obtain the upper block image;

[0025] (6.3) Take the region with a bottom height of 1 / k×h0 and a width of w0, and calculate the mean of each column to obtain a 1×w0 dimensional vector; k is the coefficient;

[0026] (6.4) Take an empty matrix of size h0×w0, and fill each row with the vector obtained in the previous step; to obtain the next block image;

[0027] (6.5) Subtract the upper block image and the lower block image from the image I0 respectively, and perform Gaussian filtering respectively to obtain the enhanced image of the upper block image and the enhanced image of the lower block image.

[0028] In step 4, for each of the enhanced images obtained in step 3—the enhanced left block image, the enhanced right block image, the enhanced upper block image, and the enhanced lower block image—multi-threshold step-by-step segmentation is performed, including the following steps:

[0029] 1) Set the number of pixels in the image that are greater than the initial threshold to 1;

[0030] 2) Find all connected regions that do not intersect the boundary, calculate the sum of the areas of all found connected regions, and push them onto the stack;

[0031] 3) Stop iteration when the current threshold is greater than the termination threshold;

[0032] Find the iteration number with the largest area in the stack, and remove all regions that intersect the boundary or have an area smaller than the set value to obtain the final segmentation result.

[0033] Step 5 involves finding the bounding rectangle based on the regions segmented by the multi-level thresholds. The steps are as follows:

[0034] (1) Calculate the sum of pixels in each column of the segmentation result;

[0035] (2) Find the column with the largest sum;

[0036] (3) Search for columns that are less than the threshold k from the left and right sides of the column, and use them as the left and right boundaries respectively;

[0037] (4) Calculate the sum of pixels in each row of the segmentation result;

[0038] (5) Find the row with the largest sum;

[0039] (6) Search for rows above and below the current row and rows less than the threshold k, and use them as the upper and lower boundaries respectively.

[0040] An infrared ship target detection system for complex sea and sky backgrounds includes:

[0041] The ship detection classifier building module is used to build the Adaboost ship detection classifier; the input image is passed through the Adaboost ship detection classifier to generate coarse ship detection boxes;

[0042] The region expansion module is used to expand the coarse detection frame region;

[0043] The region enhancement module is used to perform block-based row and column mean enhancement on the expanded coarse detection box region to obtain an enhanced image;

[0044] The image segmentation module is used to perform step-by-step image segmentation on the enhanced image;

[0045] The object detection module is used to obtain object detection results from the segmentation results.

[0046] The present invention has the following beneficial effects and advantages:

[0047] 1. Compared with deep learning methods, the algorithm designed in this invention is simpler, has lower hardware performance requirements, is easier to deploy, and has a faster inference speed.

[0048] 2. Compared with other traditional image processing detection methods, it can detect more targets, has higher positioning accuracy, and a lower false alarm rate.

[0049] 3. The block image enhancement method in this patent is more suitable for infrared ship images with complex backgrounds than the current morphological image enhancement methods, and can effectively enhance ships with low contrast on the sea surface.

[0050] 4. The step-by-step threshold segmentation algorithm designed in this patent can segment the enhanced target more effectively than the traditional threshold segmentation algorithm, with more accurate segmentation contours and fewer missegments. Attached Figure Description

[0051] Figure 1 A flowchart for infrared ship target detection against complex sea and sky backgrounds;

[0052] Figure 2 This is the result of the Adaboost test;

[0053] Figure 3 To expand the area;

[0054] Figure 4a Enhancement result for the left-side image block;

[0055] Figure 4b Enhancement result for the right-side image block;

[0056] Figure 5a The result of image enhancement for the upper block;

[0057] Figure 5bThe result of image enhancement for the next segment;

[0058] Figure 6 This is a flowchart of the step-by-step threshold segmentation process;

[0059] Figure 7 Image showing the results of threshold segmentation at each level;

[0060] Figure 8 The results are as follows.

[0061] Figure 9 For other example target detection results. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0063] like Figure 1 As shown, an infrared ship target detection method for complex sea and sky backgrounds includes the following steps:

[0064] Step 1: Build the Adaboost ship detector. Based on HOG features, train the Adaboost ship detector to generate coarse detection boxes.

[0065] Step 2: Expand the detection area output by Adaboost;

[0066] Step 3: Perform block-based row and column mean enhancement on the detection area;

[0067] Step 4: Apply stepwise image segmentation to the enhanced image;

[0068] Step 5: Obtain the target detection result through the segmentation result.

[0069] Step 1 involves training a ship detection classifier based on HOG features. An adaptive weight adjustment method is used to repeatedly train the base classifier. In each iteration, the training sample weights are adjusted based on the previous classification error rate. The specific update strategy is as follows:

[0070]

[0071] a t =log 10 [(1-ε t ) / ε t (2)

[0072] ε t =P x (h t (x i )≠y i (3)

[0073] Where r represents a correctly classified case in the previous iteration, and w represents a incorrectly classified case in the previous iteration. Z t Represents the normalization factor, a t Represents the weak classifier h t The weighting coefficients h of (x) t (x i ) indicates that the weak classifier is effective for the input sample x. i The predicted value, ε t Calculate the current weight distribution D t (x) Classification error rate on the training set (P) x (Represents probability). y i This indicates a label used to identify the type of vessel.

[0074] Increase the weights of misclassified samples to gradually strengthen the weak classifiers. Then, have all weak classifiers perform classification judgments. Finally, sum the classification results according to the error rates of the weak classifiers, and compare the weighted sum with the average classification result to obtain the classification result of the base classifier. The specific formula is as follows:

[0075]

[0076] Where T represents the total number of weak classifiers, h t Let t represent the t-th weak classifier.

[0077] A cascade method is used to combine multiple base classifiers obtained after training the Adaboost ship detection classifier in step one. In this embodiment, the parameters of the Adaboost ship detection classifier are set as follows: the number of base classifiers is 20, the base classifier recall threshold is 0.9995, and the base classifier false alarm rate threshold is 0.6. The 20 classifiers are cascaded in order of increasing complexity.

[0078] The test regions are judged sequentially. If a region is identified as a target, it is passed to the next base classifier; otherwise, it is excluded.

[0079] The Adaboost detector produces coarse detection boxes, such as Figure 2 .

[0080] In step 2, the Adaboost output detection area is expanded by 3 times in height and 2 times in width. This results in an expanded area with height h0 and width w0, as shown below. Figure 3 .

[0081] Step 3 applies Gaussian filtering to the expanded image from Step 2 to obtain I0. The image is then divided into two blocks: left-right blocks and top-bottom blocks.

[0082] For left and right blocks: (1) Take the leftmost region with height h0 and width 1 / 8w0, and calculate the mean value of each row to obtain an h0×1 dimension vector; (2) Take an empty matrix of size h0×w0, and fill each column with the vector obtained in the previous step; (3) Process the right side in the same way as the left side, and (4) Subtract the left and right block images from I0 respectively, and perform Gaussian filtering to obtain the left and right block image enhancement results respectively, such as Figure 4a , Figure 4b .

[0083] For the upper and lower blocks: (1) Take the region with the highest point being 1 / 8h0 and the width being w0, and calculate the mean value of each column to obtain a 1×w0 dimension vector; (2) Take an empty matrix of size h0×w0, and fill each row with the vector obtained in the previous step; (3) Process the lower half in the same way as the upper half; (4) Subtract the left and right block images using I0 respectively, and perform Gaussian filtering to obtain the enhancement results of the left and right block images respectively, such as Figure 5a , Figure 5b .

[0084] In step 4, the enhanced image results obtained in step 3 (left, right, top, and bottom blocks) are subjected to multi-threshold step-by-step segmentation. The steps are as follows, and the flowchart is shown below. Figure 6 .

[0085] 1) Set the initial segmentation threshold, the end threshold, and the iteration step size. Pixels in the image larger than the initial threshold are set to 1;

[0086] 2) Find all connected regions that do not intersect the boundary, calculate the sum of the areas of all found regions, and push them onto the stack;

[0087] 3) Stop iteration when the current threshold is greater than the termination threshold;

[0088] Find the iteration number with the largest area in the stack, and remove all regions that intersect the boundary or have an area less than 25 to obtain the final segmentation result, such as... Figure 7 .

[0089] Step 5, which involves segmenting the region using multi-level thresholds, requires finding the bounding rectangle. The steps are as follows:

[0090] 1) Calculate the sum of each column in the segmentation result;

[0091] 2) Find the column with the largest sum;

[0092] 3) Find columns whose sum is less than 3 from this column to the left and right, and use these columns as the left and right boundaries;

[0093] 4) Calculate the sum of each row in the segmentation result;

[0094] 5) Find the row with the largest sum;

[0095] 6) Find rows whose sum is less than 3 from the row above and below, and use these rows as the upper and lower boundaries.

[0096] Test results as follows Figure 8 The red box shows the detection results.

[0097] The designed method was validated on other infrared ship images, such as Figure 9 The method can still accurately locate the ship's position, demonstrating its robustness.

Claims

1. A method for detecting infrared ship targets against complex sea and sky backgrounds, characterized in that, Includes the following steps: Step 1: Construct the Adaboost ship detection classifier; the input image is passed through the Adaboost ship detection classifier to generate coarse ship detection boxes; Step 2: Expand the coarse detection frame area; Step 3: Perform block-based row and column mean enhancement on the expanded coarse detection box region to obtain the enhanced image; Step 4: Perform step-by-step image segmentation on the enhanced image; Step 5: Obtain the target detection result from the segmentation result; In step 1, the Adaboost ship detection classifier uses an adaptive weight adjustment method to train the base classifier. In each iteration, the weights of the training samples are adjusted according to the previous classification error rate to gradually enhance the weak classifier. In step 3, the expanded coarse detection box region is subjected to block-based row and column mean enhancement, which includes the following steps: Gaussian filtering is applied to the expanded coarse detection box region image to obtain image I0. Image I0 is then divided into left and right blocks and top and bottom blocks respectively. In step 4, for each of the enhanced images obtained in step 3—the enhanced left block image, the enhanced right block image, the enhanced upper block image, and the enhanced lower block image—multi-threshold step-by-step segmentation is performed, including the following steps: 1) Set the number of pixels in the image that are greater than the initial threshold to 1; 2) Find all connected regions that do not intersect the boundary, calculate the sum of the areas of all found connected regions, and push them onto the stack; 3) Stop iteration when the current threshold is greater than the termination threshold; Find the iteration number with the largest area in the stack, and remove all regions that intersect the boundary or have an area smaller than the set value to obtain the final segmentation result; Step 5 involves finding the bounding rectangle based on the regions segmented by multi-level thresholds. The steps are as follows: (1) Calculate the sum of pixels in each column of the segmentation result; (2) Find the column with the largest sum; (3) Search for columns that are less than the threshold k from the left and right sides of the column, and use them as the left and right boundaries respectively; (4) Calculate the sum of pixels in each row of the segmentation result; (5) Find the row with the largest sum; (6) Search for rows above and below the current row and rows that are less than the threshold k, and use them as the upper and lower boundaries respectively.

2. The infrared ship target detection method for complex sea and sky backgrounds according to claim 1, characterized in that, In step 2, the height of the coarse detection frame area is increased by m times and the width is increased by n times, resulting in an expanded area with height h0 and width w0.

3. The infrared ship target detection method for complex sea and sky backgrounds according to claim 1, characterized in that, The left and right division includes the following steps: (5.1) Take the leftmost region with height h0 and width 1 / k×w0, and calculate the mean of each row to obtain an h0×1 dimension vector; k is the coefficient; (5.2) Take an empty matrix of size h0×w0, fill each column with the vector obtained in the previous step, and obtain the left block image; (5.3) Take the rightmost region with height h0 and width 1 / k×w0, and calculate the mean of each row to obtain the h0×1 dimension vector; (5.4) Take an empty matrix of size h0×w0, fill each column with the vector obtained in the previous step, and obtain the right block image; (5.5) Subtract the left block image and the right block image from the image I0 respectively, and perform Gaussian filtering respectively to obtain the enhanced image of the left block image and the enhanced image of the right block image.

4. The infrared ship target detection method for complex sea and sky backgrounds according to claim 1, characterized in that, The upper and lower blocks include the following steps: (6.1) Take the region with a top height of 1 / k×h0 and a width of w0, and calculate the mean of each column to obtain a 1×w0 dimensional vector; k is the coefficient; (6.2) Take an empty matrix of size h0×w0, and fill each row with the vector obtained in the previous step; to obtain the upper block image; (6.3) Take the region with a bottom height of 1 / k×h0 and a width of w0, and calculate the mean of each column to obtain a 1×w0 dimensional vector; k is the coefficient; (6.4) Take an empty matrix of size h0×w0, and fill each row with the vector obtained in the previous step; to obtain the next block image; (6.5) Subtract the upper block image and the lower block image from the image I0 respectively, and perform Gaussian filtering respectively to obtain the enhanced image of the upper block image and the enhanced image of the lower block image.

5. An infrared ship target detection system for complex sea and sky backgrounds, the system being used to implement the infrared ship target detection method for complex sea and sky backgrounds as described in any one of claims 1-4, characterized in that, include: The ship detection classifier building module is used to build the Adaboost ship detection classifier; the input image is passed through the Adaboost ship detection classifier to generate coarse ship detection boxes; The region expansion module is used to expand the coarse detection frame region; The region enhancement module is used to perform block-based row and column mean enhancement on the expanded coarse detection box region to obtain an enhanced image; The image segmentation module is used to perform step-by-step image segmentation on the enhanced image; The object detection module is used to obtain object detection results from the segmentation results.

Images