A SAR Image Oil Spill Detection Method Based on Image Saliency Analysis

Abstract

A method for oil spill area detection without human interaction is proposed. This scheme is based on image saliency analysis and an adaptive iterative thresholding method for SAR image oil spill area detection. In this scheme, image saliency detection is introduced into SAR oil spill detection, and then the relationship between the adaptive iterative threshold and saliency is utilized to accurately extract the oil spill area.

Description

Technical Field

[0001] This invention relates to the fields of image processing and remote sensing imaging technology, specifically, to a method for detecting oil spills in SAR images based on image saliency analysis. Background Technology

[0002] In today's society, oil remains a vital resource. With increasingly scarce land resources and rapidly growing human demand for energy, the offshore oil industry and maritime oil transportation are booming. Marine oil spills refer to the loss of oil to varying degrees during offshore extraction or transportation. These mainly include crude oil leaks from oil wells during offshore exploration and development, leaks from near-shore pipelines or during oil tanker loading and unloading, oil spills caused by ship collisions, capsizing, or grounding, and even oil spills caused by natural disasters. These incidents all pollute the marine ecosystem to varying degrees and cause substantial economic losses.

[0003] To reduce the occurrence of oil spills, it is necessary to strengthen the monitoring and detection of marine oil spills. Synthetic Aperture Radar (SAR) has advantages such as all-weather, all-day coverage, wide range, and high precision, and has become the most effective means of oil spill monitoring. In marine oil spill detection, the spill area is an important parameter for measuring the severity of an oil spill. Therefore, accurately extracting the boundary of the oil spill area from SAR marine oil spill images is a key issue in calculating the oil spill area. In SAR marine oil spill images, the oil spill area is usually dark, showing a significant difference from the surrounding non-oil spill areas. Based on this, the oil spill area can be extracted using image processing methods.

[0004] Numerous scholars both domestically and internationally have conducted extensive and meaningful research on oil spill detection using SAR images. Liu Shanwei et al. combined the FCM and DRLSE models for SAR oil spill extraction, validating the method's effectiveness. Furthermore, they analyzed the statistical, textural, and polarimetric characteristics of fully polarimetric SAR oil spill images, establishing a complete oil spill SAR atlas through feature extraction and selection. They also introduced a multi-kernel learning method based on prior labels for oil spill detection and extraction.

[0005] Wei et al. used a single threshold segmentation method, a maximum entropy segmentation method, and an unsupervised classification method to perform target detection on SAR oil spill images. They roughly divided the images into foreground and background regions, and manually selected some oil spill areas and non-oil spill areas as regions of interest. They then statistically analyzed commonly used texture features of SAR images in the regions of interest, and combined the results of different target detections with the original images to perform classification based on a BP neural network, which yielded good results.

[0006] Yao Qi et al. conducted an oil spill extraction experiment on the sea surface near Dangan Island in the Pearl River Estuary using SAR imagery, and analyzed the different applicability of artificial neural network methods and Markov chain methods in oil spill monitoring. Duan Yaping et al. studied an oil spill detection method based on SAR imagery using deep learning methods, and realized an oil spill detection algorithm based on a combination of gray-level co-occurrence matrix and convolutional neural network.

[0007] Guo Yue et al. fused gray-level co-occurrence matrix and Tamura features to directly extract features from the original SAR images. Then, they applied the classification method of deep belief network to classify and identify three types of samples: oil film, oil film-like material, and seawater, achieving a good recognition accuracy.

[0008] Zheng Honglei et al. introduced polarization features and single scattering relative difference as characteristic parameters for oil spill detection, and developed an oil spill detection algorithm based on polarization features and artificial neural networks.

[0009] CN201210024538.7 discloses a method for segmenting oil spill images on the sea surface based on polarimetric SAR data fusion. This method first constructs an active contour energy functional based on the maximum a posteriori probability criterion for the segmented region, and represents the distribution of the segmented region as a Gibbs prior probability model. Then, the active contour model is embedded into a high-dimensional level set function, and the evolution equation is obtained using the Euler-Lagrange formula. The model includes a CFAR edge detection weighted boundary length term and a fused data statistical distance term.

[0010] CN201110277737.4 discloses a method and apparatus for detecting oil spills on the sea surface based on SAR images. The method includes: converting the SAR image into a binary image by thresholding the grayscale values ​​of pixels in the SAR image of the sea surface; determining whether the number and / or proportion of pixels with a pixel value of 1 in a neighborhood of a predetermined size for each pixel in the binary image is greater than a predetermined value; identifying neighborhoods where the number and / or proportion of pixels with a pixel value of 1 is greater than the predetermined value; and using the boundary of the image formed by all identified neighborhoods as the initial zero level set for level set segmentation to detect oil spills on the sea surface.

[0011] CN201310382104.9 discloses an oil spill detection method for complex SAR image scenes, comprising the following steps: 1. Reading in the detection image; 2. Performing image segmentation on the detection image, extracting dark spots, and processing to obtain a bright sea dark spot image; 3. In the bright sea dark spot image, setting the portion outside the dark sea region to 0 or 1 to obtain a dark sea image, performing image segmentation on the dark sea image, extracting dark spots, and processing to obtain a dark sea dark spot image; 4. Adding the dark spots from the dark sea dark spot image to the bright sea dark spot image to obtain a full dark spot image, removing false dark spots to obtain a partial dark spot image; 5. Setting a reference gray level based on the partial dark spot image, performing a miss search on the denoised detection image to obtain missed dark spots, adding the missed dark spots to the partial dark spot image to obtain the final oil spill dark spot image. This invention is applicable to oil spill dark spot extraction in complex scenes.

[0012] CN201610715334.6 discloses a method for detecting oil spills on the sea surface based on C-band polarimetric SAR imagery, including the following steps: Step 1: Radar image preprocessing; Step 2: Constructing a high-dimensional polarimetric feature set; Step 3: Constructing a linear Laplacian map dimensionality reducer to reduce dimensionality and performing k-means classification; Step 4: Using sea surface wind field data to assist in oil spill detection; Step 5: Accuracy evaluation.

[0013] CN201810373026.9 discloses a precise oil spill detection method based on CFAR. First, global CFAR is used to perform coarse detection on the area to be detected, extracting suspected oil film regions to obtain a binary reference image of the oil film target. Then, morphological filtering and other methods are used to filter the binary reference image of the oil film target to eliminate clutter interference. Finally, an adaptive window CFAR algorithm is used to perform fine detection on the filtered image, ultimately obtaining the oil film region.

[0014] CN201811463066.9 discloses a method for level set SAR oil spill extraction based on bilateral filtering. This method utilizes a bilateral filter to filter oil spill SAR images and constructs an energy function for the DRLSE model based on bilateral filtering. Energy function of DRLSE model based on bilateral filtering Energy minimization was performed; utilizing The energy minimization equation F extracted SAR oil spill information.

[0015] Given the drawbacks of level set methods requiring manual initialization and neural network methods requiring pre-calibrated samples and training to generate recognition models, designing a more accurate oil spill area extraction method based on salient images is an important research direction. Summary of the Invention

[0016] To address the problems in existing technologies, this invention provides a method for oil spill area detection without human interaction. This scheme is based on image saliency analysis and an adaptive iterative thresholding method for SAR image oil spill area detection. In this scheme, image saliency detection methods are introduced into SAR oil spill detection, and then the relationship between the adaptive iterative threshold and saliency is utilized to accurately extract the oil spill area.

[0017] This invention provides a method for detecting oil spill areas in SAR images, comprising:

[0018] Step 1) Obtain the saliency image:

[0019] Perform image saliency detection and generate a normalized saliency image;

[0020] Step 2) Calculate the adaptive iterative threshold T:

[0021] The adaptive iterative threshold T is calculated using an adaptive iterative threshold algorithm;

[0022] Step 3) Determine the oil spill area:

[0023] After performing saliency detection on the SAR oil spill image, the pixel value of each pixel in the saliency image represents the degree of saliency of that pixel. The threshold T of the entire saliency image is obtained by using an adaptive iterative threshold method. If the value of a pixel in the saliency image is greater than T, it belongs to the oil spill area; otherwise, it belongs to the non-oil spill area.

[0024] Further, step 1) involves obtaining the saliency image as follows:

[0025] 1.1) Convert the original image from RGB space to Lab space;

[0026] 1.2) Apply Gaussian filtering to the Lab space image;

[0027] 1.3) Take the mean values ​​LM, AM, and BM of the three channels of the transformed image L, a, and b respectively; and calculate and sum the Euclidean distance between the mean images of the three channels and the Gaussian filtered images respectively.

[0028] 1.4) Normalize the saliency image by using the maximum and minimum values ​​in the saliency image to generate a normalized saliency image.

[0029] Further, step 1) involves obtaining the saliency image as follows:

[0030] 1.1) First, the RGB color space of the original image I is converted to the XYZ space using equation (1), and then the XYZ space is converted to the Lab space using equation (2) to obtain the Lab space image I corresponding to the original image. Lab ;

[0031]

[0032]

[0033] in,

[0034] 1.2) Image I for Lab space Lab The three components I L I a and I b All images were filtered using a 3×3 Gaussian convolution kernel to obtain the filtered image I. GLab ;

[0035] Among them, the 3×3 Gaussian convolution kernel is

[0036] 1.3) Calculate I respectively Lab The three components I of the Lab space L I a and I b The mean values ​​LM, AM, and BM are then used to calculate the sum of these three mean values ​​and I using equation (3). GLab The Euclidean distances of the three components in the image are used to obtain the saliency image SM.

[0037] SM(x,y)=(L IGLab (x,y)-LM) 2 +(a IGLab (x,y)-AM) 2 +(b IGLab (x,y)-BM) 2 (3)

[0038] Where (x,y) are pixel coordinates, L IGLab (x,y), a IGLab (x,y) and b IGLab (x, y) are respectively I GLab The values ​​of the three components at coordinates (x, y).

[0039] 1.4) Using Equation (4), the maximum value Max in the saliency image SM is determined. SM and minimum value Min SM Normalize the SM to obtain the normalized significance image NSM.

[0040] NSM(x,y)=(SM(x,y)-Min SM ) / (Max SM -Min SM (4).

[0041] Furthermore, step 2) of the adaptive iterative threshold method is as follows:

[0042] 2.1) Find the maximum and minimum grayscale values ​​of the image, denoted as Zmax and Zmin respectively, and set the initial threshold.

[0043] T0=(Z max +Z min ) / 2 (5)

[0044] 2.2) Segment the image into foreground and background based on the threshold Tk, and calculate the average gray value Z of each. o and Z b ;

[0045] 2.3) Calculate the new threshold:

[0046] T k+1 =(Z o +Z b ) / 2 (6)

[0047] 2.4) If T k =T k+1 If the threshold is reached, the iteration stops, and the result is the final threshold; otherwise, proceed to step 2.2.

[0048] Furthermore, step 3) involves determining the oil spill area as follows:

[0049] Equation (7) is used to determine whether a pixel is an oil spill pixel by judging the relationship between the saliency value of each pixel in the normalized saliency image NSM and T.

[0050]

[0051] Where R oil (x, y) represents whether the pixel at coordinates (x, y) is a pixel in the oil spill area. A value of 1 indicates that the pixel is in the oil spill area, and a value of 0 indicates that the pixel is not in the oil spill area.

[0052] This invention combines image saliency detection methods and adaptive iterative thresholding methods and introduces them into the extraction of oil spill areas in SAR images.

[0053] First, the SAR image is processed using an image saliency detection method to generate a saliency image, making the oil spill area in the image more visually obvious. Then, the optimal segmentation threshold is obtained using an adaptive iterative thresholding method. Finally, based on the saliency image, the accurate oil spill area is extracted according to the relationship between the saliency value of each pixel and the threshold.

[0054] Experimental results show that the scheme designed in this invention can effectively extract oil spill areas in SAR images, and has high recall and accuracy.

[0055] The method designed in this invention has the following advantages:

[0056] 1) The image saliency detection method can improve the contrast between the oil spill area and the non-oil spill area, making the oil spill area and the non-oil spill area more visually distinguishable, laying the foundation for the subsequent oil spill area extraction.

[0057] 2) The adaptive iterative thresholding method can efficiently obtain the most suitable threshold for a given image, thereby better distinguishing between oil spill areas and non-oil spill areas;

[0058] 3) This invention can provide more efficient detection for large-scale marine oil spill area detection and provide a reference for manual interpretation. Attached Figure Description

[0059] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0060] Figure 1 This is SAR oil spill image example 1.

[0061] Figure 2 This is the significance test result of Example 1.

[0062] Figure 3 (a) shows the detection results of the oil spill area in Example 1.

[0063] Figure 3 (b) shows the detection results of the oil spill area in Comparative Example 1.

[0064] Figure 3 (c) shows the detection results of the oil spill area in Comparative Example 2.

[0065] Figure 3 (d) is the result of manual interpretation of Example 1.

[0066] Figure 4 This is SAR oil spill image example 2.

[0067] Figure 5 This is the significance test result of Example 2.

[0068] Figure 6 (a) shows the detection results of the oil spill area in Example 2.

[0069] Figure 6 (b) shows the detection results of the oil spill area in Comparative Example 3.

[0070] Figure 6 (c) shows the detection results of the oil spill area in Comparative Example 4.

[0071] Figure 6 (d) is the result of manual interpretation of Example 2. Detailed Implementation

[0072] The experimental data used in this paper comes from the NOWPAP (Northwest Pacific Action Plan) database.

[0073] NOWPAP covers the marine environment and coastal areas from approximately 121°E to 143°E and from approximately 33°N to 52°N.

[0074] The SAR images are C-band SAR images from the ERS-1 (European Remote Sensing Satellite-1) and ESR-2 (European Remote Sensing Satellite-2) satellites, using VV polarization. Figure 1 and Figure 4 As shown.

[0075] Example 1: A method for detecting oil spill areas in SAR images, such as SAR images... Figure 1 As shown, it includes:

[0076] Step 1) Obtain the saliency image:

[0077] 1.1) Convert the original image from RGB space to Lab space;

[0078] First, the RGB color space of the original image I is converted to the XYZ space using equation (1). Then, the XYZ space is converted to the Lab space using equation (2) to obtain the Lab space image I corresponding to the original image. Lab ;

[0079]

[0080]

[0081] in,

[0082] 1.2) Apply Gaussian filtering to the Lab space image;

[0083] Image I for Lab space Lab The three components I L I a and I b All images were filtered using a 3×3 Gaussian convolution kernel to obtain the filtered image I. GLab ;

[0084] Among them, the 3×3 Gaussian convolution kernel is

[0085] 1.3) Take the mean values ​​LM, AM, and BM for the three channels of the transformed image L, a, and b, respectively; and calculate the Euclidean distance for the mean images of the three channels and the Gaussian filtered images, respectively.

[0086] Find I respectively Lab The three components I of the Lab space L I a and I b The mean values ​​LM, AM, and BM are then used to calculate the sum of these three mean values ​​and I using equation (3). GLab The Euclidean distances of the three components in the image are used to obtain the saliency image SM.

[0087] SM(x,y)=(L IGLab (x,y)-LM) 2 +(a IGLab (x,y)-AM) 2 +(b IGLab (x,y)-BM) 2 (3)

[0088] Where (x,y) are pixel coordinates, L IGLab (x,y), a IGLab (x,y) and b IGLab (x, y) are respectively I GLab The values ​​of the three components at coordinates (x, y);

[0089] 1.4) Normalize the saliency image using the maximum and minimum values ​​in the saliency image to generate a normalized saliency image;

[0090] Using Equation (4), the maximum value Max in the saliency image SM is determined. SM and minimum value Min SM Normalize the SM to obtain the normalized significance image NSM;

[0091] NSM(x,y)=(SM(x,y)-Min SM ) / (Max SM -Min SM (4)

[0092] Perform image saliency detection and generate a normalized saliency image;

[0093] Step 2) Calculate the adaptive iterative threshold T:

[0094] The adaptive iterative threshold T is calculated using an adaptive iterative threshold algorithm;

[0095] 2.1) Find the maximum and minimum gray values ​​of the image, denoted as Z. max and Z min Let the initial threshold be

[0096] T0=(Z max +Z min ) / 2 (5)

[0097] Among them, Z max =240, Z min =22

[0098] 2.2) Segment the image into foreground and background based on the threshold Tk, and calculate the average gray value Z of each. o and Z b ;

[0099] 2.3) Calculate the new threshold:

[0100] T k+1 =(Z o +Z b ) / 2 (6)

[0101] 2.4) If T k =T k+1 If the threshold is reached, the iteration stops, and the result is the final threshold; otherwise, proceed to step 2.2.

[0102] The adaptive iterative threshold T = 135.884 for the normalized saliency image NSM;

[0103] Step 3) Determine the oil spill area:

[0104] Equation (7) is used to determine whether a pixel is an oil spill pixel by judging the relationship between the saliency value of each pixel in the normalized saliency image NSM and T.

[0105]

[0106] Where R oil (x, y) represents whether the pixel at coordinates (x, y) is a pixel in the oil spill area. A value of 1 indicates that the pixel is in the oil spill area, and a value of 0 indicates that the pixel is not in the oil spill area.

[0107] Comparative Example 1

[0108] Unlike Example 1, the level set method is used for oil spill area detection.

[0109] The level set method uses the Demo1 program in the v0 version of the level set code designed by Chunming Li. The Gaussian fuzzy variance sigma is 1.5, the smoothing Dirac function parameter epsilon is 1.5, the time step is 5, the internal energy penalty parameter mu is 0.2 / timestep, the weighted length coefficient lambda is 5, the weighted area coefficient alf is 10, and the number of iterations is 600.

[0110] Comparative Example 2

[0111] Unlike Example 1, the OTSU dynamic threshold method is used for oil spill area detection.

[0112] The OTSU dynamic threshold method is:

[0113] 1) For image I, let T be the segmentation threshold between foreground and background, the proportion of foreground points in the image be ω0, and the average gray level be μ0; the proportion of background points in the image be ω1, and the average gray level be μ1.

[0114] 2) After T traverses from the minimum gray value to the maximum gray value, take T0, which maximizes the variance δ = ω0 × ω1 × (μ0 - μ1) × (μ0 - μ1), as the optimal segmentation threshold.

[0115] The specific implementation method is to directly use the graythresh() function in MATLAB software to obtain the segmentation threshold T0 of the OTSU method.

[0116] Example 2: A method for detecting oil spill areas in SAR images, such as SAR images... Figure 4 As shown, it includes:

[0117] Step 1) Obtain the saliency image:

[0118] 1.1) Convert the original image from RGB space to Lab space;

[0119] First, the RGB color space of the original image I is converted to the XYZ space using equation (1). Then, the XYZ space is converted to the Lab space using equation (2) to obtain the Lab space image I corresponding to the original image. Lab ;

[0120]

[0121]

[0122] in,

[0123] 1.2) Apply Gaussian filtering to the Lab space image;

[0124] Image I for Lab spaceLab The three components I L I a and I b All images were filtered using a 3×3 Gaussian convolution kernel to obtain the filtered image I. GLab ;

[0125] Among them, the 3×3 Gaussian convolution kernel is

[0126] 1.3) Take the mean values ​​LM, AM, and BM for the three channels of the transformed image L, a, and b, respectively; and calculate the Euclidean distance for the mean images of the three channels and the Gaussian filtered images, respectively.

[0127] Find I respectively Lab The three components I of the Lab space L I a and I b The mean values ​​LM, AM, and BM are then used to calculate the sum of these three mean values ​​and I using equation (3). GLab The Euclidean distances of the three components in the image are used to obtain the saliency image SM.

[0128] SM(x,y)=(L IGLab (x,y)-LM) 2 +(a IGLab (x,y)-AM) 2 +(b IGLab (x,y)-BM) 2 (3)

[0129] Where (x,y) are pixel coordinates, L IGLab (x,y), a IGLab (x,y) and b IGLab (x, y) are respectively I GLab The values ​​of the three components at coordinates (x, y);

[0130] 1.4) Normalize the saliency image using the maximum and minimum values ​​in the saliency image to generate a normalized saliency image;

[0131] Using Equation (4), the maximum value Max in the saliency image SM is determined. SM and minimum value Min SM Normalize the SM to obtain the normalized significance image NSM.

[0132] NSM(x,y)=(SM(x,y)-Min SM ) / (Max SM -Min SM (4)

[0133] Perform image saliency detection and generate a normalized saliency image;

[0134] Step 2) Calculate the adaptive iterative threshold T:

[0135] The adaptive iterative threshold T is calculated using an adaptive iterative threshold algorithm;

[0136] 2.1) Find the maximum and minimum gray values ​​of the image, denoted as Z. max and Z min Let the initial threshold be set;

[0137] T0=(Z max +Z min ) / 2 (5)

[0138] In this embodiment, Zmax = 255, Zmin = 0;

[0139] 2.2) Segment the image into foreground and background based on the threshold Tk, and calculate the average gray value Z of each. o and Z b ;

[0140] 2.3) Calculate the new threshold:

[0141] T k+1 =(Z o +Z b ) / 2 (6)

[0142] 2.4) If T k =T k+1 If the threshold is reached, the iteration stops, and the result is the final threshold; otherwise, proceed to step 2.2.

[0143] The adaptive iterative threshold T = 119.0381 for the normalized saliency image NSM;

[0144] Step 3) Determine the oil spill area:

[0145] 3.1) Use the method in step 2) to obtain the adaptive iterative threshold T for the normalized saliency image NSM;

[0146] 3.2) Use equation (7) to determine whether a pixel is an oil spill pixel by judging the relationship between the saliency value of each pixel in the normalized saliency image NSM and T;

[0147]

[0148] Where R oil (x, y) represents whether the pixel at coordinates (x, y) is a pixel in the oil spill area. A value of 1 indicates that the pixel is in the oil spill area, and a value of 0 indicates that the pixel is not in the oil spill area.

[0149] Comparative Example 3

[0150] Unlike Example 2, the level set method is used for oil spill area detection.

[0151] The level set method uses the Demo1 program in the v0 version of the level set code designed by Chunming Li. The Gaussian fuzzy variance sigma is 1.5, the smoothing Dirac function parameter epsilon is 1.5, the time step is 5, the internal energy penalty parameter mu is 0.2 / timestep, the weighted length coefficient lambda is 5, the weighted area coefficient alf is 10, and the number of iterations is 600.

[0152] Comparative Example 4

[0153] Unlike Example 2, the OTSU dynamic threshold method is used for oil spill area detection.

[0154] The OTSU dynamic threshold method is:

[0155] 1) For image I, let T be the segmentation threshold between foreground and background, the proportion of foreground points in the image be ω0, and the average gray level be μ0; the proportion of background points in the image be ω1, and the average gray level be μ1.

[0156] 2) After T traverses from the minimum gray value to the maximum gray value, take T0, which maximizes the variance δ = ω0 × ω1 × (μ0 - μ1) × (μ0 - μ1), as the optimal segmentation threshold.

[0157] The specific implementation method is to directly use the graythresh() function in MATLAB software to obtain the segmentation threshold T0 of the OTSU method.

[0158] Based on manually interpreted ground truth data (i.e., manually labeled oil spill areas), the performance of three methods in oil spill area extraction was compared. The comparison metrics included recall and precision, calculated as follows:

[0159]

[0160]

[0161] The recall and accuracy of the oil spill area detection results for Examples 1-2 and Comparative Examples 1-4 are shown in Table 1.

[0162] Table 1. Recall and accuracy of oil spill area detection results for Examples 1-2 and Comparative Examples 1-4

[0163]

[0164] As shown in Table 1,

[0165] The recall rate for Example 1 was 81.04%, and the precision rate was 85.86%.

[0166] The recall rate for Comparative Example 1 was 58.37%, and the precision rate was 96.28%.

[0167] The recall rate for Comparative Example 2 was 57.12%, and the precision rate was 98.03%.

[0168] Example 2 had a recall rate of 81.79% and a precision rate of 94.48%.

[0169] The recall rate for Comparative Example 3 was 81.32%, and the precision rate was 90.97%.

[0170] The recall rate for Comparative Example 4 was 79.43%, and the precision rate was 94.31%.

[0171] Analyzing the results, we can see that:

[0172] (1) Results of oil spill area extraction: The level set method performs poorly in edge preservation, such as Figure 3 (b) and Figure 6 (b) As shown in the red circle, the OTSU method distinguishes between oil spill areas and non-oil spill areas based on a threshold, without considering the correlation between pixels within the region. Therefore, it also performs poorly in some areas, such as... Figure 3 (c) and Figure 6 (c) As shown in the green circle.

[0173] (2) Oil spill area extraction accuracy: For Figure 1 The SAR images shown, although both the level set method and the OTSU method have high accuracy, have excessively low recall; for Figure 4 The SAR images shown all demonstrate that all three methods have relatively high recall and accuracy, but the adaptive iterative thresholding method performs best.

[0174] (3) Degree of human intervention: The level set method requires manual setting of the initial contour in advance, and different initial contours and different number of iterations will lead to different results in the extraction of oil spill areas; the scheme designed in this paper does not require human intervention, nor does it require a calibrated sample training set.

[0175] It is evident that the oil spill area detection results of this invention have a recall rate and accuracy rate both exceeding 80%, demonstrating higher extraction precision.

Claims

1. A method for detecting oil spill region in SAR image without human interaction, characterized in that, The detection method comprises: Step 1) obtaining a saliency image: Carrying out image saliency detection to generate a normalized saliency image, comprising the following steps: 1.1) converting the original image from an RGB space to an Lab space; 1.2) carrying out Gaussian filtering on the Lab space image; 1.3) taking the mean values LM, AM and BM of the converted image L, a, b three channel images respectively; and calculating the square of the Euclidean distance and sum of the three channel mean value images and the Gaussian filtered images respectively, and calculating to obtain a saliency image by using the following formula: SM(x, y) = (L IGLab (x, y) - LM) 2 +(a IGLab (x, y) - AM) 2 +(b IGLab (x, y) - BM) 2 where SM is the saliency map, (x, y) is the pixel coordinate, L IGLab (x, y), a IGLab (x, y), and b IGLab (x, y) are the values of the three components of the Gaussian filtered image at coordinate (x, y), respectively; 1.4) normalizing the saliency image by using the maximum and minimum values in the saliency image to generate a normalized saliency image; Step 2) calculating an adaptive iterative threshold T: Calculating an adaptive iterative threshold T by using an adaptive iterative threshold algorithm, comprising the following steps: 2.1) Find the maximum and minimum gray values of the image, denoted as Z max and Z min , respectively, and let the initial threshold T0 = (Z max + Z min ) / 2 2.2) according to the threshold T k The image is segmented into foreground and background, and the average gray value Z o and Z b are calculated for each of them respectively. 2.3) obtaining a new threshold value: T k+1 = (Z o + Z b ) / 2 2.4) if T k = T k+1 , then stop iteration and the result is the final threshold, otherwise go to step 2.2); Step 3) judging an oil spill area: After carrying out saliency detection on the SAR oil spill image, in the saliency image, the pixel value of each pixel represents the saliency degree of the pixel; the adaptive iterative threshold method is used to obtain the threshold T of the entire saliency image, if the value of a certain pixel in the saliency image is greater than T, it belongs to the oil spill area, otherwise it belongs to the non-oil spill area.

2. The detection method according to claim 1, characterized in that: Step 1) the steps of obtaining a saliency image are as follows: 1.1) first convert the RGB color space of the original image I into XYZ space by means of formula (1), and then convert the XYZ space into Lab space by means of formula (2) to obtain the Lab space image I corresponding to the original image Lab ; wherein 1.2) the image I in the Lab space Lab of the three components I L , I a and I b are filtered with a 3x3 Gaussian kernel to obtain the filtered image I GLab ; wherein the 3x3 Gaussian convolution kernel is 1.3) Compute I Lab the mean of the three components I ab M, AM and BM of the spatial components I L , I a and I b , compute the Euclidean distances of these three means to the three components in I GLab and obtain the saliency map SM; 1.4) Normalizing SM using the maximum Max and the minimum Min in the saliency map SM of formula (4) SM SM to obtain the normalized saliency map NSM.​ NSM(x, y) = (SM(x, y) - Min SM ) / (Max SM - Min SM ) (4).

3. The detection method according to claim 1, characterized in that: Step 3) the steps of judging an oil spill area are as follows: The relationship between the saliency value of each pixel in the normalized saliency image NSM and T is determined by using formula (7) to determine whether it is an oil spill pixel; wherein R oil (x, y) represents whether a pixel at coordinate (x, y) is a pixel of the oil spill region, and is 1 if it is considered as an oil spill region pixel, and is 0 if it is considered as a non-oil spill region pixel.

Images