Two-Parameter CFAR Ship Detection Algorithm Based on Rayleigh Distribution in SAR Images

: Synthetic Aperture Radar (SAR) is an active type of microwave remote sensing. Using 1 the microwave imaging system, remote sensing monitoring of the land and global ocean can be 2 done in any weather conditions around the clock. Detection of SAR image targets is one of the main 3 needs of radar image interpretation applications. In this paper, an improved two-parameter CFAR 4 algorithm based on Rayleigh distribution and morphological processing is proposed to perform 5 ship detection and recognition in high resolution SAR images. Through simulation experiments, 6 comprehensive study of the two algorithms for high resolution SAR image target detection is 7 achieved. Finally, the results of ship detection experiments are compared and analyzed, and the 8 effects of detection are evaluated according to the Rayleigh distribution model and algorithms. 9


Introduction
SAR is a high-resolution microwave imaging radar. Compared with optical remote 12 sensing, its image is not restricted by environmental factors such as weather and light, 13 and it can detect and identify objects of interest all the time and in all weather [1][2][3][4]. 14 In military use, SAR is used primarily in the fields of target recognition, battlefield 15 surveillance, and precision targeting; in civil use, it is mainly used in fields such as 16 earth monitoring, disaster warning and marine environment monitoring [5,6]. With the 17 continuous development of SAR technology, the detection and subsequent identification 18 of ground targets (such as vehicles, ships, special buildings) have a very important value. 19 Today's technology in the field of SAR image target recognition is becoming more 20 mature. Researchers have systematically analyzed SAR target segmentation and ex- 21 plained algorithms of different technical levels, such as Constant False Alarm Rate 22 (CFAR) segmentation technology, MRF-based target segmentation and Maxflow [7]. 23 The theory of computer vision graphics is used for image segmentation and the OTSU 24 algorithm is used to achieve automatic image segmentation. For target detection, CFAR 25 algorithm, Beamlet-based SAR image target detection method, feature value extraction 26 by two-dimensional main component analysis and many other solutions have been 27 studied mainly, and they have achieved well expected results [8]. For the most impor- 28 tant part of target segmentation and target detection in SAR image target recognition, 29 researchers have developed the algorithms with the highest recognition rates, and their 30 average recognition rates have remained above 90%. CFAR is the deepest, most widely 31 used, and best-effective type of many target detection algorithms. This algorithm was ship target detection investigation [9]. Its core is to establish a distribution model based 34 on local area data, draw the probability density curve of the model, and then calculate 35 the object's pixel segmentation threshold using the false alarm rate. Finally, the target 36 with a high gray value in the SAR image is detected by threshold. To this end, this 37 paper proposes an improved two-parameter CFAR algorithm based on the Rayleigh 38 distribution model, to suppress background noise by adjusting for sea noise in the SAR 39 image. Generally, the probability density function of sea clutter and a small amount of 40 noise tends to deviate from the Rayleigh distribution, but there are still some clutter 41 backgrounds with significant non-Rayleigh characteristics. The non-Rayleigh clutter 42 background is different from the Rayleigh clutter background, containing the scale 43 and shape parameters. In the real mess environment, these two parameters are often 44 unknown, so this paper adopts the two-parameter CFAR detection algorithm. 45 Nevertheless, current research on SAR image target recognition is mainly focused 46 on performing basic algorithms and improving the recognition rate. There is no sufficient 47 research on SAR image target recognition suitable for various complex backgrounds and 48 reduction of detection target distortion. These problems need to be solved through actual 49 work. Therefore, this paper uses mathematical morphology (MM), an image analysis 50 subject based on lattice theory and topology. It is the basic theory of mathematical 51 morphology image processing and can be used in areas such as noise suppression and 52 image segmentation [10]. Its basic operations include Erosion and Dilation, Opening 53 and Closing operations. Due to factors such as an unreasonable setting of the false alarm 54 rate or an incorrect setting of the background disorder, certain false alarms will occur in 55 the target detection process. Some pixels whose sizes do not meet the target parameters 56 are also identified as targets. In these cases, morphological operations can be used to 57 remove them. For the filtered SAR image, a simple eroding operation can remove small 58 nonsensical objects in the image and speckle noises with independent high brightness 59 that is erroneously judged as the target pixel. The simple dilation operation can fill the 60 black hole caused by the low value speckle noises in the target area and the missing 61 target pixels, and can connect the adjacent unconnected target area pixels. (S3 and S6) and wide-format image [11].

73
Experimental data is collected from more than 40,000 SAR images in the SAR dataset 74 mentioned above. According to the uniform background (stable state of the sea), the 75 noise background, the land and marine background (coast or port), they are divided into 76 three groups. Several satellite images are randomly selected by each, including image 77 data from GF-3 and Sentinel-1A. Figure 1 shows some example images of the SAR dataset     [11]. The red, green, and cyan rectangles indicate the ship location, the object name, and the ship chip shape in the image on the left.

Target detection in CFAR-based SAR images
87 CFAR is a pixel-level target detection algorithm. For targets such as vehicles, ships, 88 and aircrafts, they are generally required to have a strong contrast to the background 89 clutter. Target detection is achieved by judging whether the gray value of each pixel 90 exceeds a certain preset value [12]. In this case, the detection threshold is generally 91 related to the false alarm rate, the statistical pattern of the background disorder around where x is the gray value of the detected pixel and T is the detection threshold.

95
According to Equation (1), when x > T, the detected pixel is the target itself; when 96 x ≤ T, the detected pixel is the background.   The SAR image itself features a lot of speckled noise and has a low signal-to-noise 111 ratio. Typically, prior knowledge of the background and targets is difficult to obtain, so 112 are random changes in the background environment and complex background texture 113 information, so it is very important to use statistical models for target detection. The 114 clutter statistical model for SAR images counts on statistical methods to describe SAR 115 imaging data [13]. At present, the most widely used clutter statistical models mainly 116 include the Rayleigh distribution model, the Weibull distribution model, the Pearson 117 distribution model, the Gamma distribution model, and the G 0 distribution model.

118
The Rayleigh distribution is the most common type of distribution used to describe the study of SAR image clutter statistical models, it is a multiplicative noise one which 123 has been widely used in processing, analyzing, and modeling of SAR images [14]. Its 124 probability density function is: where σ is the distribution parameter, which can be calculated using the Rayleigh 126 distribution formula shown in Equation (3).
where µ is the mean.
128 Figure 4 shows the Rayleigh probability distribution curve at different σ.

135
The purpose of CFAR radar detection is to keep the false alarm rate during the de-136 tection process at a relatively constant level. The CFAR detector traverses all pixels in the 137 complete SAR image through a sliding window to achieve the purpose of detecting SAR 138 image targets. The commonly used CFAR sliding window is a hollow one, consisting 139 of three parts: the target zone, the protection zone, and the clutter zone [15]. CFAR can 140 be divided into global and local CFAR algorithms according to window size [16]. If the 141 size is smaller than that of the SAR image, it will be a local CFAR algorithm ( developed for binary images and later expanded to functional and grayscale images [10].

169
In terms of image processing, MM is divided into binary morphology and grayscale 170 morphology, frequently applied to image segmentation, thinning, skeleton extraction, 171 edge extraction, shape analysis, corner detection, watershed algorithms, etc.

172
The object of morphological operations is a set [18]. Suppose that both the binary 178 Figure 6. Schematic diagrams of morphological erosion and dilation operations have been shown in [19]. The pink graphic shows above, erosion consists of marking the pixels at the origin of the sub-image that are the same as the structural elements of the destination image (as in subfigure (a)). The image boundary can be made to shrink to eliminate small, nonsensical objects to obtain (c). For the blue graphic, dilation consists of combining the background points in contact with the target area on the target and expanding the boundary of the target outward (as in subfigure (b)).
The function is to fill some gaps in the target area and remove the small particle noise contained in it to obtain (c). In addition, the opening and closing operations are the superposition of the two previous basic operations. The process of first erosion and then dilation is called opening, and the process of first dilation and then erosion is called closing.
Certain false alarms will occur in the target detection process, due to factors such 179 as unreasonable or incorrect setting of the false alarm rate, and complex background 180 disorder. Some pixels whose size does not meet the object parameters are also recognized

204
In the following section, the CFAR algorithm proposed in this paper will be used 205 to realize ship detection on the above SAR dataset [11], and the detection results will 206 be compared with those of the traditional two-parameter CFAR algorithm [20] for 207 evaluation research. In this paper, a large-scale remote sensing SAR dataset [11] is selected for experi-  window and background window, and these slide into the image by a certain distance.

228
The protection window is to prevent the part of the ship in the target window from 229 leaking into the background window [20]. Gets the threshold of this local window by

249
In land and sea cases, the detector also segments ship targets, but the processing of 250 ground noise appears to be the same as that of ships. In most cases, there is a lot of 251 background clutter and false alarms in CFAR detection results, which must be filtered 252 by filtering or MM to obtain detection results. Figure 9 shows the marking result of the 253 ship detection, the image targets are marked as correctly detected targets and false alarm 254 targets by red and yellow frame respectively, the blue is marked as a target whose ship 255 shape is distorted (such as broken) and counted as being within the correct detection 256 target range, and the missed targets are marked with white circles. It can be seen that the  accuracy. Across the three groups of evaluations, whether the shape of the ship target is 275 distorted also needs to be noticed. In the task of SAR target detection and recognition, it 276 is necessary to ensure that the target shape is not distorted or broken as much as possible, 277 which will help improve the detection efficiency and target recognition accuracy. The traditional two-parameter CFAR algorithm is based on the Gaussian distribu-279 tion of noise background [22]. Through the previous experimental results of Ai et al. [20],  The detection results are shown in Tables 3 and 4

298
In the above experiments, the algorithm in this paper has achieved pleasing detec-299 tion results. Whether it is the target segmentation part or the morphological filtering, 300 the ships are detected and identified. Nevertheless, in addition to physical factors such 301 as the size and shape of the ship, ocean conditions are also a critical factor affecting 302 ship detection capabilities due to the uncertainty of the swell and background noise.

303
When the sea breeze increases, the backscatter from the ocean surface increases, and 304 the brightness contrast between the ship and the ocean surface echo decreases [23,24], 305 which interferes with the algorithm's detection ability to some degree. In follow-up 306 research, the influence of different distribution patterns must be considered, and the 307 appropriate target detection algorithm needs to be investigated for each situation. In 308 addition to image noise, there are many factors such as angle of incidence, ship size, 309 wind speed, sea waves, that affect ship detectability [25,26]. Generally, high wind speeds 310 and poor marine environment will cause waves, which make ships surrounded by quite 311 a tough environment. In this paper, the simulation experiment is mainly applied to the  Currently, there are some computer vision jobs related to the detection of small targets.

335
The main component is a multi-scale function with high resolution, which can improve 336 detection accuracy [27][28][29][30]. Considering various types of ships, the distribution model 337 can also be combined with deep learning algorithms for ship detection and recognition.