A method and system for infrared image ship segmentation based on improved GLOH features

Abstract

The present invention proposes a method and system for infrared image ship segmentation based on improved GLOH features, relating to the field of image processing technology. The method comprises: acquiring an airborne infrared image containing a sky area, a sea area, ships at sea, and a coastline; extracting features from the airborne infrared image using an improved GLOH model to obtain a feature map; segmenting the airborne infrared image using an improved iterative threshold segmentation algorithm to obtain an image containing only the coastline; fusing the feature map with the image containing only the coastline to obtain a ship area image; and extracting ship targets from the ship area image to obtain a ship image. The present invention improves the accuracy of ship target extraction by deeply mining thermal information in infrared images to accurately distinguish ship targets from sea backgrounds and other interference objects. The present invention can effectively cope with various harsh environments and noise interference, significantly reducing the false recognition rate.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to an infrared image ship segmentation method and system based on improved GLOH features. Background Art

[0002] Traditional visible light monitoring technology, constrained by weather conditions (such as rain, fog, and haze) and light variations (such as the alternation of day and night and strong direct sunlight), cannot meet all-weather and all-day monitoring needs. In contrast, infrared imaging technology, with its thermal radiation sensing properties, offers all-weather capabilities, can penetrate clouds and fog, and achieve nighttime imaging, making it an indispensable core technology for maritime target monitoring.

[0003] The significant difference in thermal radiation between high-temperature components such as engines and smokestacks and the cool surrounding seawater creates a physical basis for infrared detection. Airborne infrared sensors can effectively identify ships by capturing their thermal radiation signals. However, practical applications still face the following challenges: ships typically appear as small, low-contrast targets in infrared images and are strongly interfered with by the complex background of the sea (such as waves, cloud reflections, and ocean thermal noise), resulting in extremely low discrimination between the target and the background. Therefore, there is an urgent need to develop efficient target segmentation methods to accurately extract ship targets from complex backgrounds. Summary of the Invention

[0004] In view of the above problems, the present invention proposes an infrared image ship segmentation method and system based on improved GLOH features to solve the problem of target detection in complex sea background in the existing technology.

[0005] According to one aspect of the present invention, a method for infrared image ship segmentation based on improved GLOH features is proposed, the method comprising:

[0006] Acquire airborne infrared images covering the sky, sea, ships and coastline;

[0007] Using the improved GLOH model to extract features from the airborne infrared image to obtain a feature map;

[0008] Using an improved iterative threshold segmentation algorithm to segment the airborne infrared image to obtain an image containing only the coastline;

[0009] fusing the feature map with the image containing only the coastline to obtain a ship area image;

[0010] Perform ship target extraction on the ship area image to obtain a ship image.

[0011] Furthermore, the feature extraction of the airborne infrared image using the improved GLOH model includes:

[0012] Using an improved multi-scale Harris corner detection method to detect key points in the airborne infrared image;

[0013] For the detected key points, calculate the gradient strength and direction of its neighborhood;

[0014] For any point in the neighborhood, convert its coordinates to polar coordinates relative to the key point; partition the area around the key point into polar coordinates, and count the gradient strengths in different directions in each partition;

[0015] Construct a three-dimensional histogram based on the gradient strength in different directions;

[0016] The three-dimensional histogram is flattened into a one-dimensional vector and normalized to obtain the feature descriptor, and then the feature map is obtained.

[0017] Furthermore, the detecting key points in the airborne infrared image using the improved multi-scale Harris corner detection method includes: calculating the corner point response value of each pixel point according to the following formula:

[0018]

[0019] Where, Represented in pixels (x, y) and scale The corner point response value under ; k is an empirical constant; is a matrix The determinant of is a matrix traces, is the autocorrelation matrix:

[0020]

[0021] The standard deviation is Gaussian filter kernel, represents the convolution calculation, Represents the image at scale Gaussian derivative gradients in the x-axis and y-axis directions of the lower image coordinate system;

[0022] If the corner point response value of a pixel is greater than a preset threshold, the pixel is determined to be a key point.

[0023] Furthermore, in the improved multi-scale Harris corner detection method, the original gradient calculation is improved to scale-adaptive gradient calculation, and the calculation formula is:

[0024] =

[0025] Where, Indicated on scale σ n The image gradient vector under ; Indicated on scale σ n The Gaussian smoothed image below; is the derivative operator in the x-axis direction, is the derivative operator in the y-axis direction.

[0026] Furthermore, the method of detecting key points in the airborne infrared image using the improved multi-scale Harris corner detection method further includes: further screening the key points detected based on the corner response values ​​using a gradient direction consistency formula; the gradient direction consistency formula is:

[0027]

[0028] Where, represents the standard deviation of angle, which represents the angle at multiple scales σ n The discreteness of the lower angle value θ; is the set of angles, is the set of all scales σ n The set of corresponding angle values ​​θ; represents the nth scale σ n The angle value below; represents the average value of angles at all scales; N represents the total number of scales;

[0029] If the angle standard deviation of a key point is less than the threshold , then keep the key point.

[0030] Furthermore, the formula for threshold updating in the improved iterative threshold segmentation algorithm is:

[0031]

[0032] Where, represents the dynamic threshold; is the proportionality coefficient; Represents the gradient vector of pixel point (x,y); Represents the global maximum value of the image gradient vector magnitude.

[0033] Furthermore, fusing the feature map with the image containing only the coastline to obtain the ship area image includes:

[0034] Performing maximum white pixel row detection on the image containing only the coastline to obtain the row number with the largest accumulation of white pixels;

[0035] Assigning white to the image area above the row number with the largest accumulation of white pixels in the image containing only the coastline;

[0036] The feature map is subtracted from the image after being assigned a white value, and the subtraction result is multiplied by the airborne infrared image to obtain an image of the ship area.

[0037] Furthermore, the formula for detecting the maximum white pixel row for the image containing only the coastline is:

[0038]

[0039] Where m and n represent the number of rows and columns of the image, r is the row index, and c is the column index; represents an image containing only the coastline; Indicates the row index corresponding to the maximum value.

[0040] Furthermore, a processing method based on Otsu threshold and morphology is used to extract ship targets from the ship area image to obtain a ship image, including: extracting non-zero pixel grayscale values ​​from the ship area image; then using the Otsu threshold to separate waves and ships; and then performing small area filtering to obtain a ship target image.

[0041] According to another aspect of the present invention, a ship segmentation system for infrared images based on improved GLOH features is proposed, the system comprising:

[0042] An image acquisition module configured to acquire airborne infrared images encompassing the sky, sea, marine vessels, and coastline;

[0043] a feature extraction module configured to extract features from the airborne infrared image using an improved GLOH model to obtain a corresponding feature map;

[0044] A coastline segmentation module is configured to segment the feature map using an improved iterative threshold segmentation algorithm to obtain an image containing only the coastline;

[0045] The ship extraction module is configured to fuse the feature map and the image containing only the coastline to obtain a ship area image; and perform ship extraction on the ship area image to obtain a ship image.

[0046] The beneficial technical effects of the present invention are:

[0047] In terms of target recognition accuracy, traditional methods are limited by their feature extraction capabilities, making it difficult to accurately capture subtle differences in thermal radiation from marine vessels in infrared images. In terms of environmental adaptability, existing technologies are sensitive to environmental factors such as complex sea conditions and changeable weather, making them prone to misidentification. This paper proposes a method and system for infrared image ship segmentation based on improved GLOH features. By deeply mining thermal information in infrared images, a more targeted feature description model—the improved GLOH model—is constructed. This model can accurately distinguish ship targets from the sea background and other interfering objects, significantly improving target extraction accuracy. By utilizing optimized algorithms and robust design, this paper can effectively cope with various harsh environments and noise interference, significantly reducing the misidentification rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily apparent by reading the following detailed description with reference to the accompanying drawings, in which several embodiments of the present invention are shown by way of example and not limitation, in which:

[0049] Figure 1 This is a flow chart of a method for infrared image ship segmentation based on improved GLOH features according to an embodiment of the present invention;

[0050] Figure 2 This is an example diagram of an airborne infrared image in an embodiment of the present invention;

[0051] Figure 3 2 is a flow chart of feature extraction using the improved GLOH model in an embodiment of the present invention;

[0052] Figure 4 is a feature map obtained by extracting features using the improved GLOH model in an embodiment of the present invention;

[0053] Figure 5 An image containing only the coastline is obtained using the improved iterative threshold segmentation algorithm in an embodiment of the present invention;

[0054] Figure 6 is a ship area image obtained in an embodiment of the present invention;

[0055] Figure 7 This is a diagram showing the final ship target extraction result in an embodiment of the present invention;

[0056] Figure 8 It is a structural diagram of an infrared image ship segmentation system based on improved GLOH features according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided solely to enable those skilled in the art to better understand and implement the present invention, and are not intended to limit the scope of the present invention in any way. Rather, these embodiments are provided to make this disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0058] Those skilled in the art will appreciate that embodiments of the present invention may be implemented as a system, apparatus, device, method, or computer program product. Therefore, the present disclosure may be implemented in the following forms: entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or in a combination of hardware and software. It should be understood that any number of elements in the figures is for illustrative purposes only and not limiting, and any nomenclature is for distinction only and does not have any limiting meaning.

[0059] The present invention proposes an infrared image ship segmentation method and system based on improved GLOH features to overcome the technical bottlenecks of the existing technology in marine target monitoring applications, such as the inability to effectively suppress sea surface background interference, high computational complexity, and difficulty in meeting real-time requirements when segmenting weak marine ship targets in airborne infrared images, thereby improving the segmentation effect of marine ship targets.

[0060] The embodiment of the present invention proposes a method for infrared image ship segmentation based on improved GLOH features, such as Figure 1 As shown, the method includes:

[0061] S1. Acquire airborne infrared images covering the sky, sea, ships, and coastline.

[0062] S2. extracting features from the airborne infrared image using the improved GLOH model to obtain a feature map;

[0063] S3. Segmenting the airborne infrared image using an improved iterative threshold segmentation algorithm to obtain an image containing only the coastline;

[0064] S4, fusing the feature map with the image containing only the coastline to obtain a ship area image;

[0065] S5. Extract ship targets from the ship area image to obtain a ship image.

[0066] The method starts from S1. In S1, an airborne infrared image including a sky area, a sea area, ships at sea, and a coastline is acquired.

[0067] According to an embodiment of the present invention, an airborne infrared image is an image of the sky, coastline, ocean, and ships at sea taken by a drone carrying an infrared camera in the air. Figure 2 A grayscale image of the airborne infrared image is shown.

[0068] Then, S2 is executed. In S2, the improved GLOH model is used to extract features from the airborne infrared image to obtain a feature map.

[0069] According to an embodiment of the present invention, the GLOH (Gradient Location-Orientation Histogram) model, as a local feature description method, statistically models the gradient information of local image regions by constructing gradient position and orientation histograms. This model has the following significant advantages: 1) Strong robustness: It is highly invariant to geometric transformations such as illumination changes, scale changes, and rotation, making it suitable for target feature extraction against complex ocean backgrounds; 2) Excellent anti-interference ability: Through the statistical characteristics of gradient information, it can effectively suppress ocean background noise and non-target interference, significantly improving the distinction between target and background, thereby enhancing segmentation accuracy; 3) High computational efficiency: The feature extraction process is based on local gradient statistics, with low computational complexity. Combined with optimization algorithms (such as fast gradient calculation and parallel processing), it can achieve real-time processing, meeting the stringent real-time requirements of airborne platforms.

[0070] The embodiment of the present invention improves the existing GLOH model. Figure 3 As shown in the figure, feature extraction from infrared images using the improved GLOH model involves: first, searching for key points and detecting corners in the image using the improved multi-scale Harris corner detection method. Gradient calculations are automatically adjusted based on the point scale, and the consistency of surrounding gradient directions is checked to filter out unstable points. For the remaining key points, the gradient strength and direction of the surrounding area are calculated. Partition statistics (polar coordinate partition mapping) are then performed to divide the area around the key points into multiple circular and fan-shaped blocks (polar coordinate partitioning), and the gradient strength in different directions is counted in each block. Feature concatenation (3D histogram construction) is then performed to organize these statistical results into a 3D histogram based on position and direction, which is then straightened into a large vector. Finally, normalization (normalized descriptor) adjusts this large vector to unit length and limits excessive values ​​to make the features more stable. Finally, the stable feature descriptor is output. Details are described below.

[0071] S21, perform key point detection on the infrared image, use Harris corner detector to extract key point coordinates (x, y), each key point p iAssociate a local circular area with a radius of R. In this embodiment, Harris corner detection is improved to multi-scale Harris corner detection, and the improved formula is as follows:

[0072]

[0073] Where, Represented in image pixel (x, y) and scale The corner response value under the threshold is: the larger the value, the more likely the pixel (x, y) is a corner point, and the smaller the value, the more likely it is a flat area or edge. That is, if the corner response value of a pixel is greater than the preset threshold, the pixel is determined to be a key point. k is an empirical constant (usually 0.04-0.06) used to adjust the sensitivity of corner detection. The smaller k is, the more corner points are detected, and the larger k is, the stricter the corner detection is (only strong corner points are retained). is a matrix The determinant of is calculated as follows:

[0074]

[0075] It measures the degree of local change in the gradient. The larger the value, the stronger the gradient in both the x and y directions, which may be a corner point. is a matrix The calculation formula is:

[0076]

[0077] It represents the total intensity of the gradient. The larger the value, the more likely the point is an edge or corner point. is a 2×2 autocorrelation matrix (structure tensor) calculated as:

[0078]

[0079] Represents the Gaussian derivative gradient of the image in the x-axis and y-axis directions of the image coordinate system (usually calculated using Sobel or Gaussian derivative filters); scale is the scale parameter that controls Gaussian smoothing, the larger It will make the detection more robust to noise, but may lose details. It can detect finer corners, but is sensitive to noise. The standard deviation is Gaussian filter kernel, weighted average of the squared gradient and product, suppressing noise and emphasizing the current scale The following features.

[0080] Furthermore, the original gradient calculation is replaced by scale-adaptive gradient calculation, and gradient direction consistency verification is added to the original key point detection loop step. For the key points detected based on the corner point response value, they are further screened using the gradient direction consistency formula.

[0081] The scale-adaptive gradient calculation formula is:

[0082]

[0083] Where, Indicated on scale The image gradient vector (gradient field) under is a two-dimensional vector containing the partial derivatives of the image in the x-axis and y-axis directions of the image coordinate system, which is used to calculate the edges, corners and other features of the image; Indicates that the image is at scale The Gaussian smoothed image below:

[0084]

[0085] I(x,y) is the original image, is a two-dimensional Gaussian kernel, The larger it is, the blurrier the image; Represents the convolution operation; is the derivative operator in the x-axis direction, is the derivative operator in the y-axis direction.

[0086] When using the Gaussian derivative kernel in the scale-adaptive gradient calculation formula, through the convolution derivative operator: and , the convolution kernel size varies with scale Adaptive adjustment; if a fixed operator (such as Sobel) is used, the kernel size is Not relevant, just an approximation.

[0087] The added gradient direction consistency verification formula is:

[0088]

[0089] Where, is the angular standard deviation, which represents the angle at multiple scales σ n Under this condition, the discrete degree of angle value θ is is the set of angles, is the set of all scales σ n The set of corresponding angles θ; Indicates the angle value at the nth scale, Represents the average value of the angle θ under all scales; N is the total number of scales, which refers to the scale σ involved in the calculation n The number of

[0090] If the angle standard deviation of a key point is less than the threshold , then keep the key point. As an example, , which is the critical value for determining angle consistency. If the standard deviation is less than , then the angle changes at different scales are considered stable.

[0091] In key point detection of infrared images, the gradient operator is dynamically selected according to the scale of the key points, so that it can retain more effective edges in low-resolution images and improve the stability of cross-scale feature matching; the gradient direction distribution in the neighborhood of the key points is statistically analyzed, and points with chaotic directions are eliminated, which can effectively reduce redundant key points. The retained points are mostly located at the intersection of structural edges, which can increase matching reliability; there is also an improvement in computational efficiency.

[0092] S22. Image gradient calculation: After obtaining the key point, calculate the gradient magnitude and direction of each pixel in its surrounding neighborhood.

[0093] S23, Polar coordinate partition mapping: For any point in the neighborhood, convert it to polar coordinates relative to the key point; partition the area around the key point into polar coordinates, and count the gradient strengths in different directions in each partition.

[0094] S24. Three-dimensional histogram construction: After obtaining the polar coordinates of any point in the neighborhood relative to the key point, superimpose the gradient direction statistics to construct a three-dimensional histogram, which fully describes the gradient spatial distribution of the key point neighborhood.

[0095] S25, normalized descriptor: flatten the three-dimensional histogram into a one-dimensional vector and normalize it to enhance robustness:

[0096]

[0097] Where h is the input vector, which can be a vector of any dimension. is the L2 norm (Euclidean norm) of the vector, is a smoothing term (a very small constant) used to prevent the denominator from being zero, and d is the normalized output vector. Figure 4 shown.

[0098] Then, S3 is executed. In S3, the feature map is segmented using an improved iterative threshold segmentation algorithm to obtain an image containing only the coastline.

[0099] According to an embodiment of the present invention, the calculation process of the iterative threshold segmentation algorithm is as follows:

[0100] S31, initial threshold The expression is:

[0101]

[0102] Where T0 is the average brightness of the image, which can be used for automatic exposure adjustment, binarization threshold initialization (such as the Otsu algorithm), or image enhancement. m and n represent the row and column dimensions of the image, and their sizes are automatically determined by the input image resolution. I(i,j) represents the grayscale value of the image at coordinate (i,j), which is usually [0,1] or [0,255] after normalization.

[0103] S32, dynamic threshold update:

[0104]

[0105] in, is the threshold value for the kth iteration, used to segment the foreground (target) and background of the image; α represents the foreground mean weight coefficient, which controls the segmentation bias towards high grayscale areas. Its value range is (0,1). When α>0.5, it is suitable for enhancing the extraction of high-brightness targets (suitable for coastlines in aerial photography), while α<0.5 is more suitable for dark target detection. represents the foreground mean of the k-1th iteration, represents the background mean of the k-1th iteration, which changes with the iteration number k and reflects the evolution of regional statistical characteristics.

[0106] In this embodiment, the improved dynamic threshold update formula is as follows:

[0107]

[0108] in, is the scaling factor (weight), as an example, ; The dynamic threshold is a threshold that is dynamically calculated based on the maximum value of the image gradient amplitude and is used to distinguish edges or significant features. represents the global maximum value of the image gradient amplitude, is the gradient vector of the image at point (x,y).

[0109] This dynamic threshold update formula improves adaptability. The threshold is dynamically adjusted with the maximum gradient of the image, preserving edges in high-contrast areas and suppressing noise in smooth areas. It also improves balance, taking into account both denoising and detail preservation through a fixed ratio (α=0.1), avoiding the over-smoothing or under-smoothing problems of a fixed threshold. It is also highly efficient and robust, relying only on the global maximum gradient, simple to calculate, and more robust to changes in lighting and contrast.

[0110] Use Current Threshold The image is divided into foreground and background, the grayscale mean of the two parts is calculated respectively, and the average value of the two is taken as the new threshold, which becomes the core mechanism of the feedback loop and makes the threshold approach the optimal solution.

[0111] S33, iterative termination condition:

[0112]

[0113] Where, Indicates the threshold change tolerance (in this embodiment =0.8). When it decreases, the number of iterations can be increased and the accuracy can be improved. Increasing it accelerates convergence but may stop early.

[0114] S34, binary decision:

[0115]

[0116] Among them, T final It represents the final threshold when the algorithm converges, which directly determines the accuracy of the segmentation edge; Represents the indicator function, which realizes the threshold judgment; is the input image at position Pixel value of is the value of the binary image at position (i, j). The grayscale image is converted into a binary image, and each iteration produces a temporary segmentation result for subsequent evaluation.

[0117] S35, Maximum connected area selection:

[0118]

[0119] Where C represents the set of all connected components in the binary image; |C k ∣ represents the connected area C k The number of pixels, which can quantify the regional significance and suppress noise; arg max is the parameter corresponding to the maximum value, is the final selected subset. Mark all connected areas, select the area with the largest number of pixels as the effective foreground, filter out noise and small objects, ensure that the threshold optimization targets the main target, and improve the robustness of the algorithm. The image containing only the coastline after segmentation is as follows Figure 5 shown.

[0120] Then, S4 is executed. In S4, the feature map and the image containing only the coastline are fused to obtain the ship area image.

[0121] According to an embodiment of the present invention, the process of extracting the ship area image is as follows.

[0122] S41. Detect the maximum white pixel row for the image containing only the coastline, and obtain the row number with the most accumulated white pixels:

[0123]

[0124] In the formula, m and n represent the number of rows and columns of the image, The row number with the most accumulated white pixels is the final result (target row number); Represents an image containing only coastlines, r is the row index, c is the column index, represents the total number of coastline pixels in row r, The above formula locates the key horizontal area in the image by counting the pixel density of each row. It then sums the pixels of each row vertically and finds the row with the maximum cumulative pixel value. This allows us to identify the main vertical distribution of the coastline, assuming that ships do not appear above the coastline.

[0125] S42, assigning a value of white to the image area above the row number with the largest accumulation of white pixels in the image containing only the coastline;

[0126] S43, subtract the feature map from the image after assigning white value, and multiply the subtraction result with the airborne infrared image to obtain the ship area image. Figure 6 shown.

[0127] Then, S5 is executed. In S5, a ship extraction is performed on the ship area image using a processing method based on Otsu threshold and morphology to obtain a ship image.

[0128] According to an embodiment of the present invention, the process of extracting ships from a ship region image using a processing method based on Otsu threshold and morphology is as follows.

[0129] S51, non-zero pixel gray value extraction:

[0130]

[0131] in, is the input image, Represents a set of non-zero pixel values. This formula can exclude background black pixels (value 0) and only analyze the grayscale distribution of the ship area. are pixel coordinates, is the pixel value of the image at coordinate (i, j), is a collection This step removes pure black background pixels, retaining only "valid pixels" that may contain ships. Filtering out background pixels and focusing on valid pixels is often used for sparse data processing or target feature extraction.

[0132] S52, Otsu threshold calculation:

[0133]

[0134]

[0135] Where, represents the optimal threshold, T is the candidate threshold, arg max is the parameter corresponding to the maximum value, is the between-class variance, is the percentage of background pixels, is the proportion of foreground pixels, is the average grayscale of the background class, Represents the average grayscale of the foreground class, is the overall average grayscale of the image. This step achieves adaptive binarization by maximizing the separation between foreground and background using statistical methods. The optimal threshold for distinguishing between ships and waves is found. The optimal segmentation point is automatically calculated by maximizing the inter-class variance.

[0136] S53, Image Binarization:

[0137]

[0138] Among them, T actual Determined adaptively by the Otsu algorithm, the implicit assumption in this formula is that the grayscale of the ship area is significantly higher than the background (this needs to be ensured by preliminary processing). represents the pixel value of the input image at coordinate (i, j), Represents the value of the binary image at (i, j). The purpose of binarization is to convert the grayscale image into an image containing only black and white, and to simplify the image through threshold segmentation to highlight key features.

[0139] S54, small area filtering:

[0140]

[0141] Where, represents the minimum retention area threshold, as an example, A min =10 pixels; represents the input binary image, Indicates the area opening operation, including: detecting all connected regions in a binary image; calculating the number of pixels in each connected region; deleting pixels with an area smaller than area; Represents the processed binary image.

[0142] Through adaptive threshold segmentation and morphological optimization, refined extraction of the ship area is achieved, which is suitable for subsequent target recognition and tracking tasks. Figure 7 This is the final ship target extraction result after removing noise and small area interference.

[0143] Another embodiment of the present invention provides an infrared image ship segmentation system based on improved GLOH features, such as Figure 8 As shown, the system includes:

[0144] An image acquisition module 810 is configured to acquire airborne infrared images including the sky, sea, ships at sea, and coastline;

[0145] a feature extraction module 820 configured to extract features from the airborne infrared image using the improved GLOH model to obtain a corresponding feature map;

[0146] A coastline segmentation module 830 is configured to segment the feature map using an improved iterative threshold segmentation algorithm to obtain an image containing only the coastline;

[0147] The ship extraction module 840 is configured to fuse the feature map and the image containing only the coastline to obtain a ship area image; and perform ship extraction on the ship area image to obtain a ship image.

[0148] It should be noted that the functions of the infrared image ship segmentation system based on improved GLOH features described in this embodiment can be described by the aforementioned infrared image ship segmentation method based on improved GLOH features. For the parts not described in detail in the system embodiment, please refer to the above method embodiment.

[0149] It should be noted that although several units, modules, or submodules are mentioned in the detailed description above, such division is merely exemplary and not mandatory. In fact, according to embodiments of the present invention, the features and functions of two or more modules described above may be embodied in one module. Conversely, the features and functions of one module described above may be further divided and embodied by multiple modules.

[0150] Furthermore, although the operations of the method of the present invention are described in a particular order in the accompanying drawings, this does not require or imply that these operations must be performed in this particular order, or that all illustrated operations must be performed to achieve the desired results. Additionally or alternatively, some steps may be omitted, multiple steps may be combined into one step, and / or one step may be decomposed into multiple steps.

[0151] Although the spirit and principles of the present invention have been described with reference to several specific embodiments, it should be understood that the present invention is not limited to the specific embodiments disclosed, and the division into various aspects does not mean that the features of these aspects cannot be combined to benefit. Such division is only for the convenience of expression. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A ship segmentation method for infrared images based on improved GLOH features, characterized in that: include: Acquire airborne infrared images covering the sky, sea, ships and coastline; The improved GLOH model is used to extract features from the airborne infrared image to obtain a feature map. The method comprises: detecting key points in the airborne infrared image using an improved multi-scale Harris corner detection method; calculating the gradient strength and direction of the detected key points in their neighborhood; converting the coordinates of any point in the neighborhood into polar coordinates relative to the key point; partitioning the area around the key point into polar coordinates, and counting the gradient strengths in different directions in each partition; constructing a three-dimensional histogram based on the gradient strengths in different directions; flattening the three-dimensional histogram into a one-dimensional vector and normalizing it to obtain a feature descriptor, thereby obtaining a feature map. Using an improved iterative threshold segmentation algorithm to segment the airborne infrared image to obtain an image containing only the coastline; fusing the feature map with the image containing only the coastline to obtain a ship area image; Perform ship target extraction on the ship area image to obtain a ship image.

2. The infrared image ship segmentation method based on improved GLOH feature according to claim 1, characterized in that: The detecting key points in the airborne infrared image using the improved multi-scale Harris corner detection method includes: calculating the corner point response value of each pixel point according to the following formula: ； Where, Represented in pixels (x, y) and scale The corner point response value under ; k is an empirical constant; is a matrix The determinant of is a matrix traces, is the autocorrelation matrix: ； The standard deviation is Gaussian filter kernel, represents the convolution calculation, Represents the image at scale Gaussian derivative gradients in the x-axis and y-axis directions of the lower image coordinate system; If the corner point response value of a pixel is greater than a preset threshold, the pixel is determined to be a key point.

3. The infrared image ship segmentation method based on improved GLOH feature according to claim 2, characterized in that: In the improved multi-scale Harris corner detection method, the original gradient calculation is improved to scale-adaptive gradient calculation, and the calculation formula is: = ； Where, Indicated on scale σ n The image gradient vector under ; Indicated on scale σ n The Gaussian smoothed image below; is the derivative operator in the x-axis direction, is the derivative operator in the y-axis direction.

4. The infrared image ship segmentation method based on improved GLOH features according to claim 3, characterized in that: The method of detecting key points in the airborne infrared image using the improved multi-scale Harris corner detection method further includes: further screening the key points detected based on the corner response values ​​using a gradient direction consistency formula; the gradient direction consistency formula is: ； Where, represents the standard deviation of angle, which represents the angle at multiple scales σ n The discreteness of the lower angle value θ; is the set of angles, is the set of all scales σ n The set of corresponding angle values ​​θ; represents the nth scale σ n The angle value below; represents the average value of angles at all scales; N represents the total number of scales; If the angle standard deviation of a key point is less than the threshold , then keep the key point.

5. The infrared image ship segmentation method based on improved GLOH feature according to claim 1, characterized in that: The formula for threshold update in the improved iterative threshold segmentation algorithm is: ； Where, represents the dynamic threshold; is the proportionality coefficient; Represents the gradient vector of pixel point (x,y); Represents the global maximum value of the image gradient vector magnitude.

6. The infrared image ship segmentation method based on improved GLOH features according to claim 1, characterized in that: The step of fusing the feature map with the image containing only the coastline to obtain the image of the ship area includes: Performing maximum white pixel row detection on the image containing only the coastline to obtain the row number with the largest accumulation of white pixels; Assigning white to the image area above the row number with the largest accumulation of white pixels in the image containing only the coastline; The feature map is subtracted from the image after being assigned a white value, and the subtraction result is multiplied by the airborne infrared image to obtain an image of the ship area.

7. The infrared image ship segmentation method based on improved GLOH features according to claim 6, characterized in that: The formula for detecting the maximum white pixel row for the image containing only the coastline is: ； Where m and n represent the number of rows and columns of the image, r is the row index, and c is the column index; represents an image containing only the coastline; Indicates the row index corresponding to the maximum value.

8. The infrared image ship segmentation method based on improved GLOH features according to claim 1, characterized in that: The ship target is extracted from the ship area image using a processing method based on Otsu threshold and morphology to obtain a ship image, including: extracting non-zero pixel grayscale values ​​from the ship area image; then using Otsu threshold to separate waves and ships; and then performing small area filtering to obtain a ship target image.

9. An infrared image ship segmentation system based on improved GLOH features, characterized in that: include: An image acquisition module configured to acquire airborne infrared images encompassing the sky, sea, marine vessels, and coastline; A feature extraction module is configured to extract features from the airborne infrared image using an improved GLOH model to obtain a corresponding feature map. The module comprises: detecting key points in the airborne infrared image using an improved multi-scale Harris corner detection method; calculating the gradient strength and direction of the detected key points in their neighborhood; converting the coordinates of any point in the neighborhood into polar coordinates relative to the key point; partitioning the area around the key point into polar coordinates and counting the gradient strengths in different directions in each partition; constructing a three-dimensional histogram based on the gradient strengths in different directions; flattening the three-dimensional histogram into a one-dimensional vector and normalizing it to obtain a feature descriptor, thereby obtaining a feature map. A coastline segmentation module is configured to segment the feature map using an improved iterative threshold segmentation algorithm to obtain an image containing only the coastline; The ship extraction module is configured to fuse the feature map and the image containing only the coastline to obtain a ship area image; and perform ship extraction on the ship area image to obtain a ship image.

Images