A sea ship target saliency detection method imitating a raptor visual perception enhancement mechanism

Abstract

The application discloses a kind of offshore ship target saliency detection methods of simulating raptor visual perception enhancement mechanism, step one: simulating raptor visual light perception enhancement mechanism modeling;Step two: simulating raptor visual foreground perception enhancement mechanism modeling;Step three: image filtering operation;Step four: simulating raptor visual contrast perception enhancement mechanism modeling;Step five: simulating raptor visual perception equalization mechanism modeling;Step six: calculate image saliency value and obtain saliency map;Step seven: output the saliency map S of simulating raptor visual perception enhancement mechanism offshore ship target.The application introduces raptor visual mechanism into offshore target saliency detection process, can better map the ability of raptor capture long-distance prey in complex environment to the task demand of long-distance identification and tracking offshore ship target in unmanned aerial vehicle;The application uses perception enhancement mechanism to enhance the adaptability of unmanned aerial vehicle to environment, so that unmanned aerial vehicle can accurately identify and safely recover in various complex sea surface environment.

Description

Technical Field

[0001] This invention provides a method for detecting the salience of maritime targets by mimicking the visual perception enhancement mechanism of birds of prey, especially for the long-range recovery of maritime unmanned aerial vehicles (UAVs), and belongs to the field of computer vision. Background Technology

[0002] In recent years, with the depletion of land resources, countries have turned their attention to the vast oceans, which are rich in natural resources. Developing marine equipment has become a crucial task for many nations, and unmanned marine equipment is an important component of this sector. Whether for civilian or military use, unmanned marine equipment holds significant value in practical applications. Compared to manned equipment, unmanned equipment offers advantages such as no personnel casualties, lower cost, lighter weight, greater stealth capabilities, longer patrol and combat operations, and reduced personnel and manufacturing costs. Unmanned equipment is also more tolerant of harsh environments than humans, allowing it to perform more complex tasks more flexibly, reach inaccessible areas, and save on the manpower and communication resources required for traditional manned operations.

[0003] However, in practical applications, the limited field of view and slow maneuverability restrict the unmanned surface vessels (USVs) from expanding into a wider range of missions. The maneuverability and flexibility of unmanned aerial vehicles (UAVs) can effectively compensate for this deficiency, and their onboard reconnaissance and communication equipment can significantly enhance and expand the mission capabilities of USVs. Therefore, heterogeneous collaborative operations between USVs and UAVs have become an important model for future marine unmanned equipment applications, and the safe recovery of UAVs will be a crucial part of this process.

[0004] The safe recovery of unmanned aerial vehicles (UAVs) at sea relies on several key technologies, such as UAV anti-interference control technology, landing area positioning, and UAV guidance control technology. Landing area positioning can be achieved through various methods, including GPS navigation, inertial navigation, and ultrasonic positioning. However, these traditional navigation technologies are susceptible to interference from external electromagnetic signals, and their sensors are prone to accumulating errors, making real-time and accurate positioning of the landing area difficult. Visual navigation, on the other hand, employs a passive approach, requiring simple equipment, offering strong environmental adaptability, and having a wide range of applications. When UAVs are at medium to long distances at sea, unmanned surface vessels (USVs) are non-cooperative targets within the UAV's field of vision, resulting in low image resolution and limited information. As the UAV approaches the USV, the image becomes clearer, and the field of vision becomes richer, leading to increased environmental interference. At this point, the UAV needs to accurately pinpoint the specific location of the landing area in complex sea conditions to prepare for a precise landing. Bionic vision, a new and important branch of computer vision, offers a novel navigation method that is highly adaptable to weather conditions, does not require numerous external sensors or satellite navigation systems, exhibits strong anti-interference capabilities and environmental adaptability, and has significant development potential. The development of biomimetic vision technology will provide new ideas for solving the problem of maritime target recognition in complex sea conditions.

[0005] Birds of prey possess extremely high visual sensitivity, with a wide field of vision, long-range vision, and strong zoom capabilities. They excel at locating prey on the ground from high altitudes, earning them the title of "Kings of the Sky." The retina of raptors is covered with abundant photoreceptor cells, mainly divided into two categories: cone cells and rod cells. These two types are sensitive to bright and dim environments, respectively. During a hunt, the brightness and contrast of the environment change in real time. Thanks to the abundant photoreceptor cells, raptors can adapt to these changes, adjusting the activity of their cone and rod cells to navigate complex environments and capture detailed information. Furthermore, the biconical cells in the raptor's retina have a higher topological order than single-conical cells, exhibiting a "hyperhomocylindrical" distribution. This allows them to play a significant role in non-color vision, giving them an advantage in motion detection and detail detection. Studies have shown that low contrast between prey and background, as well as a relatively dense background layer, reduces the number of hunting attempts and successes for raptors. Other research indicates that, in terms of contrast sensitivity, the raptor's visual system simultaneously possesses a relatively narrow spatial frequency bandwidth and relatively high visual acuity. Therefore, contrast information is also a crucial element in the visual system of raptors, a key feature affecting target discrimination. Raptors also possess a foveal structure; the foveal region of the raptor's retina has a relatively high density of photoreceptors, enabling "fine sampling" of the imaging area and resulting in high-resolution images. In contrast, the density of photoreceptors in the peripheral region is relatively low, leading to coarser information acquisition and lower resolution. This results in different information sampling outcomes between the foveal and peripheral regions during raptor vision imaging, reflecting both differences in scene contrast and preserving more comprehensive detail. A comparison of the density and imaging results of foveal and peripheral cells is shown below. Figure 2 As shown. These unique physiological structures of birds of prey help them enhance their perception of prey in different environments, demonstrating superior hunting abilities. Considering that the detection of maritime targets is often affected by factors such as lighting, water mist, and the sea-sky boundary, the sensory enhancement mechanism of birds of prey is applied to the salience detection of maritime targets.

[0006] In summary, this invention proposes a method for detecting the salience of maritime vessels based on a raptor-inspired visual perception enhancement mechanism. The aim is to provide a real-time, online method for detecting medium- to long-range maritime vessels in complex sea environments, thereby effectively improving the detection efficiency and accuracy of maritime vessels in complex sea environments and making it possible to achieve safe recovery of UAVs at medium to long distances. Summary of the Invention

[0007] 1. Purpose of the invention:

[0008] This invention addresses the problem of medium- and long-range recovery of unmanned aerial vehicles (UAVs) at sea by proposing a method for detecting the salience of maritime vessels based on a raptor-inspired visual perception enhancement mechanism. The aim is to provide a real-time, online method for detecting medium- and long-range maritime vessels in complex sea environments, thereby effectively improving the detection efficiency and accuracy of maritime vessels in complex sea environments and making it possible to achieve safe medium- and long-range recovery of UAVs.

[0009] 2. Technical Solution:

[0010] This invention addresses the problem of identifying medium- to long-range maritime targets using unmanned aerial vehicles (UAVs). It develops a method for detecting the salience of maritime targets that mimics the visual perception enhancement mechanism of birds of prey. The framework of this method is as follows: Figure 3 As shown, the input image is first preprocessed by selecting the channel with more prominent details based on grayscale information; then, the image is enhanced based on morphological operations such as erosion, reconstruction, and dilation to highlight foreground information; then, the image is low-pass filtered to remove interference; then, the image is contrast-enhanced and equalized to highlight the target area; finally, the saliency value of the image is calculated and the saliency map is output.

[0011] The specific steps for implementing the maritime target salience detection method based on the raptor-inspired visual perception enhancement mechanism are as follows:

[0012] Step 1: Modeling the mechanism for enhancing visual illumination perception in raptors

[0013] The retinas of birds of prey are covered with abundant photoreceptor cells, mainly divided into two categories: cone cells and rod cells. These two types are sensitive to bright and dim environments, respectively. Leveraging this rich array of photoreceptor cells, birds of prey can adjust the ratio of cone cells to rod cells to enhance their perception of different lighting conditions, thus adapting to their environment and acquiring more detailed information.

[0014] The pixel value range of raptor vision imaging is finite; the highest value is imaged as white, and the lowest value as black. Grayscale reflects the brightness value between white and black; the smaller the grayscale value, the smaller the corresponding pixel brightness value. Therefore, a raptor-like vision illumination perception enhancement mechanism can be described based on grayscale information. This mechanism involves three channels: a bright channel, a dark channel, and a grayscale channel, which can be defined as follows:

[0015]

[0016] Among them, I L I D and I G These are the bright channel, dark channel, and grayscale channel of image I. r I g and I bThese are the color value channels of the red, green, and blue primary colors of image I, respectively. (x,y) are the coordinates of a pixel in image I, and Ω(x,y,r0) represents the neighborhood patch of pixel (x,y) with radius r0.

[0017] Based on the distribution and mechanism of photoreceptor cells in the raptor's retina, this invention proposes a raptor-inspired visual illumination enhancement mechanism. Specifically, when the grayscale value of a pixel is less than the average grayscale value of the image by a threshold, the rod cells in the raptor's retina activate, assigning the pixel value using the dark channel; when the grayscale value of a pixel is greater than the average grayscale value of the image by a threshold, the cone cells in the raptor's retina activate, assigning the pixel value using the bright channel; otherwise, the pixel value is assigned using the grayscale channel. This processing enhances the differentiation of illumination in visual field imaging, making dark areas darker and bright areas brighter, reflecting the mechanism of illumination enhancement. The mathematical model of the raptor-inspired visual illumination enhancement mechanism can be summarized as Equation (2).

[0018]

[0019] Among them, I LPEtemp It is an image obtained through light sensitivity enhancement adjustment, T G It is the threshold corresponding to the channel selection, mid Ggray Let I be the average gray value of the input image I. M and N are the height and width of the input image I, respectively.

[0020] For ease of subsequent processing, the complement of image A is defined as... and For image I LPEtemp Ashing can be represented as:

[0021]

[0022] Among them, I LPE T is the image after the input image I has been adjusted using a raptor-inspired visual illumination perception enhancement mechanism. gray The threshold for graying processing, For image I LPEtemp The supplement, count G For image I LPEtemp The mid-gray value is less than the average gray value mid LPEgary The number of pixels, mid LPEgary Image I LPEtemp The average grayscale value. G The calculation formula is shown in equation (4).

[0023]

[0024] Step 2: Modeling the foreground perception enhancement mechanism in raptor-like vision

[0025] Biconical cells, composed of closely packed pairs of primary and secondary cones, are the mediators of non-color vision in birds of prey. These cells are distributed in a "hyperhomocylindrical" pattern in the retina, a highly ordered distribution that makes birds of prey extremely sensitive to visual details. This study simulates the role of biconical cells in detail detection by modeling a foreground perception enhancement mechanism in simulated birds of prey based on morphological operations such as erosion, reconstruction, and dilation.

[0026] S21, Corrosion Operation

[0027] Image I adjusted in step one LPE The purpose of performing the erosion operation is to eliminate the boundary points of the image, causing the image to shrink inward along the boundary, as shown in equation (5).

[0028]

[0029] Among them, I erode It is an image obtained by erosion operation, where Se(x,y) represents the neighborhood structure of pixel (x,y).

[0030] S22, Reconstruction Operation

[0031] For the eroded image I erode Reconstruction is performed, with the aim of extracting connected regions containing labeled features from the template image to construct a new image. Based on the structuring element Se, image I... LPE Reconstructed Image I erode The transformation is denoted as I rec1 =Rc(I erode ,I LPE The transformation is defined by the iterative equation (6).

[0032]

[0033] Where h1 is initialized to flag I erode Iterate until h k+1 =h k End, at this point there is I rec1 =h k+1 .

[0034] S23, Expansion Operation

[0035] For the reconstructed image I rec1 The purpose of performing the dilation operation is to expand the boundary points of the image, which can be expressed as Equation (8).

[0036]

[0037] Among them, Idilate It is an image obtained after dilation.

[0038] S24, Based on the structural element Se, using images Reconstructed Image Image I was obtained rec2 , can be represented as:

[0039]

[0040] in, and They are I dilate and I rec1 The supplement.

[0041] The erosion-dilation operation is used to remove image regions that cannot contain neighborhood structures, smooth image contours, break narrow links, and remove small outliers, thus helping to highlight the target region from the background. In summary, we can obtain an image I adjusted using a raptor-inspired foreground perception enhancement mechanism. FPE As shown in equation (10).

[0042]

[0043] in, isI rec2 The supplement.

[0044] Step 3: Image Filtering Operation

[0045] To further filter out noise interference in the high-frequency part of the image, frequency domain filtering can be performed on the image. This invention proposes a quadratic arctangent low-pass filter, which can filter out or significantly attenuate the high-frequency components of the image (such as image noise) and smooth the image. In addition, the effective information of the image is concentrated in the low frequency, and retaining the low frequency will better preserve the details and facilitate filtering and observation. The transfer function of the quadratic arctangent low-pass filter is shown in Equation (11), and its three-dimensional image is shown in Equation (11). Figure 4 As shown.

[0046]

[0047] Where D(u,v) is the distance from pixel (u,v) to the center of the image, and D0 is the cutoff frequency.

[0048] Image I adjusted in step two FPETo perform frequency domain filtering, the image first needs to be transformed to the frequency domain through Fourier transform and the spectrum centered, as shown in Equation (12). The purpose of centering is to place the low frequencies in the center of the spectrum, concentrating the energy for easier filtering. Then, a second arctangent low-pass filter is performed on the image, as shown in Equation (13). Finally, the filtered result is subjected to inverse Fourier transform and spectrum decentering, and the real part is taken, thus obtaining the result of the second arctangent low-pass filter, as shown in Equation (14).

[0049]

[0050]

[0051]

[0052] Where F is the image I FPE The image obtained by converting to the frequency domain, where (u,v) are the pixel coordinates in the frequency domain, G is the image obtained after a second arctangent low-pass filter, and I... atanLP It is the image obtained by filtering and converting it to the time domain.

[0053] Step 4: Modeling the mechanism for enhancing visual contrast perception in raptors

[0054] Contrast information is crucial for the visual system of birds of prey, representing a key feature influencing target differentiation. Studies have shown that low contrast between prey and background, along with relatively dense background overlays, reduces the number of hunting attempts and successes in birds of prey. Further research indicates that the visual system of birds of prey possesses both a relatively narrow spatial frequency bandwidth and relatively high visual sharpness in terms of contrast sensitivity. Enhancing contrast aims to make white areas brighter and black areas darker, resulting in a more vivid image with clearer outlines and details.

[0055] This invention proposes a piecewise gray-level nonlinear stretching method based on line segment transformation and elliptic transformation to enhance image contrast. When the input gray-level value is less than a certain threshold, the gray-level value of that pixel is assigned 0; when the input gray-level value is greater than the threshold, a new gray-level value is calculated based on the elliptic transformation. The transformation of the input gray-level value and the output gray-level value is as follows: Figure 5 As shown.

[0056] So, for image I adjusted in step three atanLP Piecewise grayscale nonlinear stretching based on line segment transformation and ellipse transformation is performed, as shown in Equation (15).

[0057]

[0058] Among them, I CPE The image is obtained by adjusting the visual contrast perception enhancement mechanism similar to that of birds of prey.max and G min Image I atanLP Maximum and minimum gray values, T C It is the segmented threshold.

[0059] Step 5: Modeling the visual perception balance mechanism of raptors

[0060] The foveal region of a raptor's retina has a relatively high density of photoreceptors, enabling "fine sampling" of the imaging area and resulting in a high local resolution imaging effect. In contrast, the photoreceptor density in the peripheral region is relatively low, leading to a coarser acquisition of information and thus a relatively low resolution. This results in different information sampling results between the foveal and peripheral regions during raptor vision imaging, reflecting both the difference in scene contrast and preserving more comprehensive detail information. Based on this, a mathematical model of the raptor-like visual contrast balancing mechanism is constructed, which can be summarized as Equation (16).

[0061] I PB =(k PB ·I CPE +I atanLP ) / (k PB +1) (16)

[0062] Among them, I PB The image is obtained by adjusting the visual perception balance mechanism similar to that of birds of prey, k PB It is the proportional coefficient of the perceived equilibrium mechanism.

[0063] Step 6: Calculate the saliency value of the image and obtain the saliency map.

[0064] Image I PB The image is switched to the frequency domain and subjected to a Gaussian high-pass filter, then switched back to the time domain to obtain the filtered image S. t Then, image I CB and image S t By subtracting the values, the saliency values ​​of the images can be calculated. Finally, the images obtained by subtracting are normalized to obtain the image saliency map S.

[0065] S61. Applying image I in the frequency domain PB The Gaussian high-pass filter is applied, and the processing procedure is shown below.

[0066]

[0067]

[0068]

[0069] Among them, S f Image I PBThe image obtained by converting to the frequency domain and performing spectral centering, S HP It is an image obtained after Gaussian high-pass filtering, S t It is the image obtained by converting back to the time domain and performing spectral decentering. H Gauss It is the transfer function of the Gaussian high-pass filter, and its expression is shown in equation (20).

[0070]

[0071] Among them, D Gauss (u,v) is the distance from pixel (u,v) to the center point of the image, D Gauss0 It is the cutoff frequency.

[0072] S62, Transfer image I CB and image S t The difference is calculated, and the result is normalized to obtain the saliency plot S, as shown below.

[0073] S a =I PB -S t (twenty one)

[0074]

[0075] Among them, S a It is the result of subtracting the images, and S is the saliency map.

[0076] Step 7: Output the salience map S of maritime ship targets using the raptor-inspired visual perception enhancement mechanism.

[0077] Through calculations in steps one through six, the final saliency detection results for the UAV in identifying maritime targets in complex sea conditions are obtained. Using triangulation, further position estimation can be performed, guiding the UAV towards the maritime target and ultimately achieving safe recovery of the UAV.

[0078] 3. Advantages and effects:

[0079] This invention discloses a method for salience detection of maritime targets based on a raptor-inspired visual perception enhancement mechanism. This method selects channels rich in detail information based on grayscale information and further highlights the detailed information of the target area based on foreground perception enhancement, contrast enhancement, and equalization mechanisms. Furthermore, it reduces noise interference through two layers of filtering to achieve salience detection of maritime targets in complex sea environments. The advantages and effects of this invention are: 1) Introducing a raptor-inspired visual mechanism into the maritime target salience detection process better maps the ability of raptors to capture mid-to-long-range prey in complex environments to the task requirements of UAVs for mid-to-long-range identification and tracking of maritime targets; 2) Using a perception enhancement mechanism enhances the UAV's adaptability to the environment, enabling the UAV to accurately identify and safely recover targets in various complex sea environments. Attached Figure Description

[0080] Figure 1 Comparison of density and imaging results between central foveal cells and peripheral cells

[0081] Figure 2 Framework diagram of a maritime target salience detection method based on raptor-inspired visual perception enhancement mechanism Figure 3 3D plot of the transfer function of a second arctangent low-pass filter

[0082] Figure 4 Piecewise gray-level nonlinear stretching transformation image based on line segment transformation and elliptic transformation

[0083] Figure 5 Original image of a ship at sea under clear weather conditions

[0084] Figure 6 Image of the salience detection results of ships at sea under clear weather conditions

[0085] Figure 7 Original image of a ship at sea under cloudy conditions

[0086] Figure 8 Image of the salience detection results of ships at sea under cloudy conditions

[0087] Figure 9 Original image of a ship at sea in foggy conditions

[0088] Figure 10 Image of the salience detection results of ships at sea in foggy conditions

[0089] Figure 11 Original image of a ship at sea under twilight conditions

[0090] Figure 12 Image of the salience detection results of ships at sea under twilight conditions

[0091] Figure 13 Original image of a ship target at sea in a sea ice environment

[0092] Figure 14 Image of the salience detection results for ships at sea in sea ice environment

[0093] Figure 15 Original image of the unmanned surface vessel captured by a drone.

[0094] Figure 16 Image of target salience detection results taken by drone

[0095] The labels and symbols in the diagram are explained as follows:

[0096] I - Input Image

[0097] mid Ggray —The average gray value of input image I

[0098] T G ——Channel selection threshold

[0099] I LPEtemp —Image obtained by adjusting the visual illumination perception enhancement mechanism similar to that of birds of prey

[0100] mid LPEgray —Input image I LSEtemp Average gray value

[0101] count G —Image I LSEtemp Gray value less than mid LPEgray Number of pixels

[0102] T gray Image graying threshold

[0103] I LPE —Image obtained after graying and adjustment using a raptor-inspired visual light perception enhancement mechanism.

[0104] I FPE —Image obtained by adjusting the foreground perception enhancement mechanism similar to that of raptors

[0105] I atanLP —The image obtained after two arctangent low-pass filters

[0106] G min —Image I atanLP minimum gray value

[0107] G max —Image I atanLP Maximum grayscale value

[0108] T C Image grayscale stretching segmentation threshold

[0109] I CPE —Images adjusted using a raptor-inspired visual contrast enhancement mechanism

[0110] I PB —Images adjusted using a raptor-inspired visual perception equalization mechanism

[0111] S—Saliency map output of input image I Detailed Implementation

[0112] The effectiveness of the proposed method is verified below through a specific example of ship target salience detection in a complex marine environment. The experimental computer was configured with an Intel Core i7-6700 processor (3.40GHz), 16GB of RAM, and MATLAB R2020a software. The specific steps of a method for ship target salience detection at sea based on a raptor-inspired visual perception enhancement mechanism are as follows:

[0113] Step 1: Modeling the mechanism for enhancing visual illumination perception in raptors

[0114] The pixel value range of raptor vision imaging is finite, with the highest value corresponding to white and the lowest value corresponding to black. Grayscale value is the brightness value between white and black; the smaller the brightness value, the smaller the corresponding grayscale value. Therefore, a raptor-like vision illumination enhancement mechanism can be described based on grayscale information. This mechanism involves three channels: a bright channel, a dark channel, and a grayscale channel, defined as shown in equation (1), where I L I D and I G These are the bright channel, dark channel, and grayscale channel of image I. r I g and I b These are the color value channels of the red, green, and blue primary colors of image I, respectively. (x,y) are the coordinates of a pixel in image I, and Ω(x,y,r0) represents the neighborhood patch of pixel (x,y) with radius r0.

[0115] Based on the distribution and mechanism of photoreceptor cells in the retina of raptors, this invention proposes a raptor-inspired visual illumination enhancement mechanism. Specifically, when the gray value of a pixel is less than the average gray value of the image, the raptor's rod cells activate, assigning a value to that pixel using the dark channel; conical cells activate, assigning a value to that pixel using the bright channel. This enhances the differentiation of illumination in visual field imaging, making dark areas darker and bright areas brighter, reflecting the mechanism of illumination enhancement. The mathematical model of the raptor-inspired visual illumination enhancement mechanism can be summarized as equation (2), where I... LPEtemp It is an image obtained through light sensitivity enhancement adjustment, midGgray Let I be the average gray value of the input image I. M and N are the height and width of the input image I, respectively.

[0116] For ease of subsequent processing, the complement of image A is defined as... and For image I LPEtemp After graying, an image I is obtained after being adjusted by a raptor-inspired visual illumination perception enhancement mechanism. LPE As shown in equation (3), where T G The threshold for ashing was determined after multiple experiments, and T was selected as the threshold. G =0.5. For image I LPEtemp The supplement, count G For image I LPEtemp The mid-gray value is less than the average gray value mid LPEgray The number of pixels is calculated as shown in equation (4).

[0117] Step 2: Modeling the foreground perception enhancement mechanism in raptor-like vision

[0118] The role of bipyramidal cells in detail detection is simulated, and the mechanism of visual foreground perception enhancement in raptors is modeled based on morphological operations of erosion, reconstruction, and dilation.

[0119] S21. Adjust the image I after step one. LPE Image I was obtained by performing an etch operation. erode The neighborhood structure Se(x,y) can be constructed using the strel function in MATLAB, specifically called as Se = strel(shape, parameters), where shape is a string specifying the desired shape, and parameters are a series of parameters specifying shape information. After several experiments, shape = 'disk' was selected. The erosion operation can be implemented using the imerode function in MATLAB, specifically called as I... erode =imerode(I LPE ,Se).

[0120] S22, with image I LSE For the eroded image I erode Reconstruction is performed to obtain image I rec1 Image reconstruction can be achieved using the `imreconstruct` function in MATLAB, whose specific calling command is I. rec1 =imreconstruct(I erode ,I LSE ).

[0121] S23, For the reconstructed image I rec1 Perform a dilation operation to obtain image I. dilate The erosion operation can be implemented using the `imdilate` function in MATLAB, specifically by calling the command `I...`. dilate =imdilate(I rec1 ,Se).

[0122] S24, Image I rec1 The complement of the dilated image I dilate The image I is reconstructed by the addition of the missing parts. rec2 The specific command to rebuild is

[0123] In summary, we can obtain image I adjusted by the raptor-inspired visual foreground perception enhancement mechanism. FPE As shown in equation (10), where isI rec2 The supplement.

[0124] Step 3: Image Filtering Operation

[0125] To further filter out noise interference in the high-frequency part of the image, frequency domain filtering can be performed on the image. The transfer function of a quadratic arctangent low-pass filter proposed in this invention is shown in equation (11), where D(u,v) is the distance from pixel (u,v) to the center point of the image, and D0 is the cutoff frequency. After multiple experiments, D0 = 40 was selected.

[0126] Image I adjusted by the raptor-inspired visual foreground perception enhancement mechanism in step two PPE Frequency domain filtering is performed. First, the image needs to be transformed to the frequency domain through Fourier transform and the spectrum centered, as shown in Equation (12). Then, a second arctangent low-pass filter is performed on the image, as shown in Equation (13). Finally, the filtered result is subjected to inverse Fourier transform and spectrum centered, and the real part is taken. This gives the image I obtained after adjustment by the second arctangent low-pass filter. atanLP As shown in equation (14).

[0127] Step 4: Modeling the mechanism for enhancing visual contrast perception in raptors

[0128] This invention proposes a piecewise grayscale nonlinear stretching method based on line segment transformation and elliptic transformation, as shown in equation (15), where I CPE The image is obtained by adjusting the visual contrast perception enhancement mechanism similar to that of birds of prey. max and G min Image I atanLP Maximum and minimum gray values, T C It is the segmentation threshold. After multiple experiments, T was selected. C=20+G min .

[0129] Step 5: Modeling the visual perception balance mechanism of raptors

[0130] The foveal region of a raptor's retina has a relatively high density of photoreceptors, enabling "fine sampling" of the imaging area and resulting in a high local resolution imaging effect. In contrast, the photoreceptor density in the peripheral region is relatively low, leading to a coarser acquisition of information and thus a relatively low resolution. This results in different information sampling results between the foveal and peripheral regions during raptor vision imaging, reflecting both the difference in scene contrast and preserving more comprehensive detail information. Based on this, a mathematical model of the raptor-like visual contrast balancing mechanism is constructed, which can be summarized as Equation (16), where I... PB The image is obtained by adjusting the visual perception balance mechanism similar to that of birds of prey, k PB It is the proportionality coefficient of the equilibrium mechanism. k was selected after multiple experiments. PB =10.

[0131] Step 6: Calculate the saliency value of the image and obtain the saliency map.

[0132] Image I PB The image is switched to the frequency domain and subjected to a Gaussian high-pass filter, then switched back to the time domain to obtain the filtered image S. t Then, image I CB and image S t By subtracting the values, the saliency values ​​of the images can be calculated. Finally, the images obtained by subtracting are normalized to obtain the image saliency map S.

[0133] S61. Applying image I in the frequency domain PB Gaussian high-pass filtering is performed, and the processing procedure is shown in equations (17)-(19), where S f Image I PB The image obtained by converting to the frequency domain and performing spectral centering, S HP It is an image obtained after Gaussian high-pass filtering, S t It is the image obtained by converting back to the time domain and performing spectral decentering. H Gauss D is the transfer function of a Gaussian high-pass filter, and its expression is shown in equation (20), where D Gauss (u,v) is the distance from pixel (u,v) to the center point of the image, D Gauss0 It is the cutoff frequency, and D was selected after multiple experiments. Gauss0 =2.

[0134] S62, Transfer image I CB and image S tThe difference is calculated, and the result is normalized to obtain the saliency map S, which can be expressed as equations (21) and (22), where S a It is the result of subtracting the images, and S is the saliency map.

[0135] Step 7: Output the salience map S of maritime ship targets using the raptor-inspired visual perception enhancement mechanism.

[0136] Through calculations in steps one through six, the final saliency detection results for the UAV in identifying maritime targets in complex sea conditions are obtained. Using triangulation, further position estimation can be performed, guiding the UAV towards the maritime target and ultimately achieving safe recovery of the UAV.

[0137] Original image of a ship at sea under clear weather conditions Figure 5 As shown, the detection results of the salience of ships at sea under clear weather conditions are as follows: Figure 6 As shown, under clear weather conditions, this method can filter out the effects of clouds and waves, effectively detecting maritime targets.

[0138] Original image of a ship at sea under cloudy conditions Figure 7 As shown, the detection results of the salience of ships at sea under cloudy conditions are as follows: Figure 8 As shown, this method can effectively eliminate the influence of low brightness in overcast weather, effectively detect maritime targets, and can detect multiple targets.

[0139] Original image of a ship at sea in foggy conditions Figure 9 As shown, the detection results of the salience of ships at sea in foggy conditions are as follows: Figure 10 As shown, this method can effectively eliminate the effects of low contrast and detect maritime targets in foggy weather.

[0140] Original image of a ship at sea under twilight conditions Figure 11 As shown, the results of salience detection of ships at sea in twilight conditions are as follows: Figure 12 As shown, this method can effectively detect maritime targets at dusk by eliminating the effects of excessively high contrast and vibrant colors.

[0141] Original image of a ship target at sea in a sea ice environment (as shown) Figure 13 As shown, the detection results of the salience of ships at sea in a sea ice environment are as follows: Figure 14 As shown, this method can eliminate the interference of broken, dense sea ice in a sea ice environment and effectively detect maritime targets from a complex sea surface background.

[0142] Original image of the unmanned surface vessel captured by a drone, as shown below. Figure 15As shown, the target salience detection results of the drone collected are as follows: Figure 16 As shown in the figure, this method can effectively detect unmanned surface vessel targets.

Claims

1. A method for detecting the conspicuity of a marine vessel target by imitating the enhanced mechanism of raptor vision, characterized in that: The method comprises the following steps: Step one: modeling of the light perception enhancement mechanism of the falcon vision The light perception enhancement mechanism of the falcon vision is described based on the gray information, which involves three channels, i.e., a bright channel, a dark channel and a gray channel, and can be defined as follows: (1) wherein, , and are the light channel, dark channel and gray channel of the input image respectively, , and are the red, green and blue color value channels of the image respectively, is the coordinate of the pixel of the image, denotes the neighborhood patch of the pixel with radius . The mathematical model of the light perception enhancement mechanism of the falcon vision is summarized as formula (2): (2) wherein, is an image obtained by adjusting the image through light perception enhancement, is a threshold value corresponding to the channel selection, is an average gray value of the input image , , and are height and width of the input image , respectively. For the convenience of subsequent processing, define the complement of image as , and ; the image is subjected to graying processing, which can be represented as: (3) in, For the input image Image adjusted using a raptor-inspired visual illumination enhancement mechanism. The threshold for graying. For image The supplement, For image The medium gray value is less than the average gray value The number of pixels, Representing an image Average gray value, ; The calculation formula is shown in equation (4): (4) Step two: modeling of the foreground perception enhancement mechanism of the falcon vision The foreground perception enhancement mechanism of the falcon vision is modeled based on the erosion operation, reconstruction and dilation operation in morphology; The reconstruction of the image is based on the structure The reconstruction of the image is based on the structure The transformation is defined by the iteration (5): （5） wherein, initialized to an image ; iterate until end, at which time there are ; Then, the image adjusted in step one The erosion, reconstruction, dilation and secondary reconstruction operations are performed in the following order: (6) (7) (8) (9) (10) wherein, is the eroded image, denotes the neighborhood structure of a pixel , is the reconstructed image, is the dilated image, is the twice reconstructed image, , and are the complements of , and respectively, is the image adjusted by the simulation of the raptor visual foreground perception enhancement mechanism.

2. The method of detecting the conspicuity of a target of a marine vessel by an enhanced mechanism of falcon vision imitation according to claim 1, characterized in that: The method further comprises: Step three: image filtering operation A twice inverse tangent low-pass filter is proposed, and the transfer function of the twice inverse tangent low-pass filter is shown as formula (11): (11) wherein is a pixel point distance to the image center point, is a cut-off frequency; The image filtering operation specifically comprises: Firstly, the image is converted to the frequency domain through Fourier transform and the spectrum is centralized, then the image is subjected to twice inverse tangent low-pass filtering, finally the spectrum is decentered and the image is converted back to the time domain through inverse Fourier transform; the whole process is shown as follows: (12) (13) (14) wherein, is an image is an image obtained by converting the image is a coordinate of a pixel point of the image in the frequency domain, is an image obtained by inverse tangent low-pass filtering the image is an image obtained by converting the filtered image to the time domain.

3. The method of detecting the conspicuity of a target of a marine vessel by an artificial falcon vision feeling enhancement mechanism according to claim 2, characterized in that: The method further comprises: Step four: modeling of the contrast perception enhancement mechanism of the falcon vision A piecewise gray nonlinear stretching method based on line segment transformation and ellipse transformation is proposed to simulate the contrast enhancement of the image in the visual field by the visual system of the raptor. When the input gray value is less than a certain threshold, the gray value of the pixel is assigned 0. When the input gray value is greater than the threshold, the new gray assignment is calculated according to the ellipse transformation. Then, the image adjusted in step three is stretched by the piecewise gray nonlinear stretching based on line segment transformation and ellipse transformation as shown in equation (15). A piecewise gray nonlinear stretching method based on line segment transformation and ellipse transformation is proposed to simulate the contrast enhancement of the image in the visual field by the visual system of the raptor. When the input gray value is less than a certain threshold, the gray value of the pixel is assigned 0. When the input gray value is greater than the threshold, the new gray assignment is calculated according to the ellipse transformation. Then, the image adjusted in step three is stretched by the piecewise gray nonlinear stretching based on line segment transformation and ellipse transformation as shown in equation (15). (15) wherein, is an image adjusted by an enhanced mechanism of contrast sensitivity of falcon vision imitation, and are the maximum and minimum gray values of the image respectively, is a segmented threshold.

4. The method of detecting the conspicuity of a target of a marine vessel by an artificial falcon vision feeling enhancement mechanism according to claim 3, characterized in that: The method further comprises: Step five: modeling of the perception equalization mechanism of the falcon vision The contrast perception equalization mechanism of the falcon vision is mathematically modeled, and the model can be summarized as formula (16): (16) wherein, is an image adjusted by the falcon vision perception equalization mechanism, is a proportional coefficient of the perception equalization mechanism; Step six: calculation of the image saliency value and obtaining of a saliency map Convert the image to the frequency domain and perform a Gaussian high-pass filtering process, and then convert back to the time domain to obtain a filtered image Convert the image to the frequency domain and perform a Gaussian high-pass filtering process, and then convert back to the time domain to obtain a filtered image ; then, subtract the image from the image to calculate the image saliency value; finally, normalize the image obtained by the subtraction to obtain the image saliency map : (17) (18) (19) (20) (21) (22) wherein, is a transfer function of a Gaussian high-pass filter, is a pixel point distance to the image center point, is a cut-off frequency; is an image is an image obtained by converting to the frequency domain and performing spectral centering processing, is an image obtained by Gaussian high-pass filtering, is an image obtained by converting back to the time domain and performing spectral decentering processing; is a result of image subtraction, is a saliency map.

5. The method of detecting the conspicuity of a target of a marine vessel by an artificial falcon vision feeling enhancement mechanism according to claim 4, characterized in that: The method further comprises: Step seven: outputting a saliency map of the maritime vessel target with the enhanced mechanism of falcon vision simulation ; Through the calculation of steps one to six, the saliency detection result of the sea ship target of the complex sea surface condition recognized by the unmanned aerial vehicle is obtained; according to the manner of triangular positioning, the position estimation can be further performed, and then the unmanned aerial vehicle is guided to fly to the sea ship target, so as to finally realize the safe recovery of the unmanned aerial vehicle.

Images