Polarization weak target enhanced detection method based on bionic optic neural pathway

Abstract

The invention discloses a polarization weak target enhanced detection method based on a bionic optic neural pathway, which is used for low-altitude slow and small target detection and comprises the following steps: acquiring a polarization image by using a polarization image sensor; segmenting the static area and the dynamic area by using a frame difference method, wherein the obtained dynamic area is an area where a dynamic weak target is located in the image; calculating the polarization degree of the dynamic region image by adopting a dense polarization resolving method to obtain a polarization degree image of the dynamic region; establishing a visual enhancement model by using the bionic visual neural pathway, inputting the polarization degree image into the model, and outputting an enhanced image; and carrying out connectivity analysis on the enhanced image to obtain the position of the weak target in the airspace. According to the invention, effective enhanced detection of the weak target in different light environments such as strong light and dark light can be realized, and the detection precision of the weak target in the intrusion airspace is effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of weak target enhancement and target detection technology, specifically relating to a polarization-based weak target enhancement and detection method based on a biomimetic visual neural pathway. Background Technology

[0002] In unauthorized areas, such as airports (posing a potential risk to aircraft takeoff and landing), borders (used for illegal activities such as drug smuggling), and conflict zones (carrying explosives that could harm civilians), there is a growing number of unregulated micro-drones. Effective detection of low-altitude, weakly dynamic targets has become a key technology for ensuring airspace security. Currently, low-altitude dynamic target detection mainly relies on radar and infrared detection; however, these systems are expensive and have poor imaging quality. Low-altitude, weakly dynamic targets, such as drones, slow-moving small aircraft, and even low-flying birds, have extremely small radar cross-sections (RCS) due to their low altitude, small size, and slow speed, making accurate detection and tracking by radar systems difficult. Furthermore, the radar reflections of low-altitude, weakly dynamic targets are easily masked by ground clutter, further limiting the accuracy of radar system detection and tracking. Traditional visual detection methods based on infrared cameras suffer from poor image clarity and susceptibility to environmental influences. While most algorithms employ deep learning, data on weak aerial targets is currently difficult to collect, resulting in limited data samples and poor detection results. Therefore, traditional radar and infrared detection methods cannot accurately detect small, low-altitude targets. Summary of the Invention

[0003] Purpose of the invention: In order to solve the problems of poor imaging quality, great susceptibility to environmental influences, and low accuracy of detection results in the detection of weak targets at low altitudes in the existing technology, the present invention provides a polarization-enhanced detection method based on a biomimetic visual neural pathway.

[0004] Technical solution: A method for enhancing the detection of weak polarized targets based on biomimetic visual neural pathways, comprising the following steps:

[0005] (1) Acquire polarization images using a polarization image sensor;

[0006] (2) Use the frame difference method to separate the static region from the dynamic region. The obtained dynamic region is the region where the weak dynamic target is located in the image.

[0007] (3) The polarization degree of the dynamic region image is calculated using the dense polarization solution method to obtain the polarization degree image of the dynamic region;

[0008] (4) A visual enhancement model is established using a biomimetic visual neural pathway. The visual enhancement model includes photon receptor cells and neurons. The photon receptor cells sequentially include adaptive low-pass filtering and exponential low-pass filtering. The filtered image I weber As input to the neuron, the neuron is used to perform Naka-Rushton transformation based on image features to achieve image enhancement; the polarization degree image obtained in step (3) is input into the visual enhancement model to output the enhanced image;

[0009] (5) Perform connectivity analysis on the enhanced image to obtain the location of weak targets in the airspace.

[0010] Furthermore, in step (4), adaptive low-pass filtering refers to changing the filtering degree according to the signal strength by dynamically adjusting the cutoff frequency of the low-pass filter and the variable gain control.

[0011] Furthermore, in step (4), the specific operation of the adaptive low-pass filtering is as follows:

[0012]

[0013] In the formula I L1 For the image after adaptive low-pass filtering, f c1 f represents the angular frequency of this stage. r Indicates frame rate, I DOP The input is a polarization degree image;

[0014] The output of the low-pass filter is normalized, and the Naka-Rushton transform is used to estimate the adaptive state I. nr , means as follows:

[0015]

[0016] I mid1 It is the median intensity of the polarization images selected from the empirical dataset;

[0017] Adaptive low-pass filtering is expressed as:

[0018] I L2 =f LPF (I DOP f fm )

[0019] In the formula f fm It is the adaptive focusing frequency, and the calculation formula is:

[0020] f fm =(f max -f min )I nr +f min

[0021] f max and f min These represent the highest and lowest adaptive rates of the adaptive low-pass filter, respectively.

[0022] Finally, a nonlinear adaptive gain is used to compress the dynamic range, resulting in the image I after the adaptive low-pass filtering stage. af The formula is as follows:

[0023] I af =g af I L2

[0024] Gain factor g af The calculation is as follows:

[0025] g af =(g max -1)(1-L nr )+1

[0026] Among them, g max This is the maximum gain factor.

[0027] Furthermore, in step (4), the exponential low-pass filtering process specifically includes:

[0028] The low-pass filter circuit is represented as follows:

[0029] I L3 =f LPF (I af f c3 )

[0030] In the formula f c3 I represents the angular frequency of this stage. L3 The output image is generated by an exponential low-pass filter;

[0031] Recalculate the exponential low-pass filter result I weber :

[0032]

[0033] In the formula, α represents the sensitivity of the system.

[0034] Furthermore, in step (4), the Naka-Rushton transform simulates differential processing of signals with different intensities by adjusting the relationship between the output and the input. The adjustment formula is as follows:

[0035]

[0036] In the formula I out Indicates the response value; I represents the intensity of the stimulus light; I mid2 This indicates the intensity of a half-saturated stimulus.

[0037] Furthermore, in step (1), the polarization image sensor is a four-quadrant polarization camera.

[0038] Furthermore, in step (2), the three-frame difference method is used to extract the dynamic region in order to increase the proportion of weak targets in the image.

[0039] Compared with existing technologies, the present invention provides a method for enhancing the detection of weak polarized targets based on biomimetic visual neural pathways, which has at least the following advantages:

[0040] (1) The visual enhancement model proposed in this method is inspired by the hoverfly's sensitive perception of weak targets. It simulates the hoverfly's visual system by adjusting signal processing according to ambient light intensity through adaptive low-pass filtering, which can highlight useful signals in complex environments. It also imitates the biological mechanism by which the hoverfly gradually adapts to and maintains visual stability under different lighting conditions through exponential low-pass filtering. This effectively improves the signal-to-noise ratio and helps to improve the detection distance and accuracy of weak targets.

[0041] (2) The Naka-Rushton transformation stage in the visual enhancement model simulates the differential processing of signals of different intensities through static nonlinear transformation. This mimics the nonlinear transformation performed by the visual system of the hoverfly when processing visual signals in different brightness regions, enhancing the perception of contrast and detail. This makes the model more sensitive to weak signals in low-light environments, while avoiding over-response in strong light environments. It improves the model's ability to identify weak target signals. The system is less affected by environmental factors and can effectively enhance weak targets under different lighting conditions, such as strong light and dim light, effectively improving the detection accuracy of weak targets in intruding into the airspace.

[0042] (3) By combining polarization images and frame difference method to extract the global detection algorithm of moving target region, the calculation time of subsequent local enhancement detection algorithm is greatly reduced, thereby shortening the detection time and improving the response speed. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the dense polarization solution method in this invention;

[0044] Figure 2 This is a flowchart of a method for enhancing the detection of weak polarized targets based on biomimetic visual neural pathways;

[0045] Figure 3 The results are obtained by using a biomimetic visual neural pathway-based polarization-enhanced detection method for weak targets under different lighting conditions.

[0046] Figure 4The image shows the results of detecting weak targets at different heights using a polarization-based weak target enhancement detection method based on biomimetic visual neural pathways.

[0047] Figure 5 This is a comparison of the polarization-based weak target enhancement detection method based on biomimetic visual neural pathways with other image enhancement perception methods. Detailed Implementation

[0048] The present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments.

[0049] A biomimetic visual neural pathway-based method for enhancing the detection of weak polarized targets, such as... Figure 2 As shown, it includes the following steps:

[0050] (1) A polarization image is acquired using a polarization image sensor, denoted as I. pol The polarization image sensor is preferably a four-quadrant polarization camera. Since the transmission directions of adjacent four quadrants are different during the image acquisition process, the polarization grating has more obvious edge information when acquiring different boundaries compared to a visible light camera, which is beneficial for subsequent image segmentation.

[0051] (2) The static region and the dynamic region are separated using the frame difference method. The resulting dynamic region is the area where the weak dynamic target is located in the image. The dynamic region is magnified in the polarized image, and this part of the image is extracted. This part of the image is represented as I. in .

[0052] When detecting highly dynamic small targets, since the target occupies a small number of pixels in the entire image, using a global processing method would waste a lot of time. Therefore, a global detector + local detector method is chosen for weak target detection. First, the moving object is enhanced by segmenting the static and dynamic regions and then analyzed. Based on the characteristics of the target motion, the range of the weak target is extracted using the three-frame difference method, thereby increasing the proportion of the weak target in the image. Furthermore, since the polarization image obtained by the four-quadrant polarization camera in step (1) has more obvious edge information, the inter-frame difference method can be used to better obtain the region where the moving object is located, thereby realizing the segmentation of the dynamic and static regions in the image.

[0053] (3) The polarization degree of the dynamic region image is calculated using a dense polarization resolution method to obtain a polarization degree image of the dynamic region. This ensures that the calculated polarization degree image has the same scale as the original image, thus better preserving the detailed information of weak targets. When using polarization degree data for target detection, the reflectivity of weak targets is usually relatively stable, differing from the reflectivity of interfering targets such as clouds. This allows us to effectively separate weak targets from background interference, thereby achieving accurate detection of specific targets. In this step, a dense resolution method is chosen to calculate the polarization degree image, ensuring that the calculated polarization degree image has the same scale as the original image, thus better preserving the information of weak targets.

[0054] The extracted image I in The process of performing dense polarization solution is as follows Figure 1 As shown, the polarization degree image I of this image is obtained. DOP The specific calculations are as follows:

[0055]

[0056] (4) Inspired by the four-stage biological perception model, a visual enhancement model is established using a biomimetic visual neural pathway. The visual enhancement model includes photon receptor cells and neurons. The photon receptor cells sequentially include an adaptive low-pass filtering stage and an exponential low-pass filtering stage, which filters the image I... weber As input to the neuron, the neuron is used to perform Naka-Rushton transformation based on image features to achieve image enhancement; the polarization degree image obtained in step (3) is input into the visual enhancement model to output the enhanced image.

[0057] 1) Adaptive low-pass filtering stage

[0058] In the early stages of the hoverfly's visual system, photoreceptor cells (PRCs) are responsible for adjusting the dynamic range and processing the input signal. Similarly, in the adaptive low-pass filtering stage of the visual enhancement model, the filtering degree is changed according to the signal strength by dynamically adjusting the cutoff frequency and variable gain control of the low-pass filter. In visual images, signal power is related to spatial frequency, and noise is essentially white noise. Adaptive filtering removes high-frequency noise and improves the signal-to-noise ratio by setting an appropriate frequency threshold; this is similar to how the hoverfly's visual system adjusts signal processing according to ambient light intensity, both aimed at highlighting useful signals in complex environments.

[0059] The specific operation of adaptive low-pass filtering is as follows:

[0060]

[0061] In the formula I L1 For the image after adaptive low-pass filtering, f c1f represents the angular frequency of this stage. r Indicates frame rate, I DOP The input is a polarization degree image;

[0062] The output of the low-pass filter is normalized, and the Naka-Rushton transform is used to estimate the adaptive state I. nr , means as follows:

[0063]

[0064] I mid1 It is the median intensity of the polarization images selected from the empirical dataset;

[0065] Adaptive low-pass filtering is expressed as:

[0066] I L2 =f LPF (I DOP f fm (5)

[0067] In the formula f fm It is the adaptive focusing frequency, and the calculation formula is:

[0068] f fm =(f max -f min )I nr +f min (6)

[0069] f max and f min These represent the highest and lowest adaptive rates of the adaptive low-pass filter, which can be empirically set based on the noise levels of the sensors in the system and the noise levels of the environment.

[0070] Finally, a nonlinear adaptive gain is used to compress the dynamic range, resulting in the image I after the adaptive low-pass filtering stage. af The formula is as follows:

[0071] I af =g af I L2 (7)

[0072] Gain factor g af The calculation is as follows:

[0073] g af =(g max -1)(1-L nr )+1 (8)

[0074] Among them, g max This is the maximum gain factor.

[0075] 2) Low-pass filtering (Weber stage)

[0076] The exponential low-pass filter (Weber) stage in the visual enhancement model simulates this feedback mechanism. The Weber stage achieves long-term, slow adaptation through exponential operations and a low cutoff frequency, maintaining temporal consistency and resisting high-frequency changes, mimicking the biological mechanism by which hoverflies gradually adapt and maintain visual stability under different lighting conditions. The low-pass filter loop in this stage is represented as follows:

[0077] I L3 =f LPF (I af f c3 (9)

[0078] In the formula f c3 I represents the angular frequency of this stage. L3 The Weber stage output image is low-pass filtered.

[0079] Recalculate the exponential low-pass filter result I weber :

[0080]

[0081] In the formula, α represents the sensitivity of the system.

[0082] 3) Naka-Rushton transformation stage

[0083] In the visual system of hoverflies, the processing of visual signals involves complex nonlinear mechanisms to adapt to different light intensities and scene changes. The Naka-Rushton stage simulates the differential processing of signals of different intensities through static nonlinear transformations. Similarly, the hoverfly's visual system also performs nonlinear transformations when processing visual signals in different brightness regions to enhance the perception of contrast and detail. In low-light environments, the visual system is more sensitive to weak signals, while in strong light environments, it avoids over-response. The Naka-Rushton stage, by adjusting the relationship between output and input, simulates the differential processing of signals of different intensities, improving the model's ability to recognize weak target signals. The specific adjustment formula is as follows:

[0084]

[0085] In the formula I out Indicates the response value; I represents the intensity of the stimulus light; I mid2 Represents the half-saturated stimulus intensity, with parameters n and I in the model. mid2 The ambient light intensity and target light intensity in the experimental data can be obtained by collecting experimental data under different conditions and then experimentally fitting the data.

[0086] (5) Perform connectivity analysis on the enhanced image to obtain the location of weak targets in the airspace.

[0087] The effectiveness of this method is verified through comparative experiments. The weak target enhancement perception system is built based on the LUCIDPHX050S-P / Q polarization camera and Jetson Xavier NX, and the weak target detected is a DJI Mavic 2. Figure 3 The image shows the enhancement results of this invention under different lighting conditions. Figure 4 This is a diagram showing the detection results of the present invention at different altitudes of the drone. Figure 5 This image shows the results of comparing the present invention with other augmented perception algorithms. The algorithms used in the comparison include: Tophat algorithm, MaxMedian algorithm, Local Contrast Method (LCM), and High-Enhancement Multi-Scale Local Contrast Method (HB-MLCM). Figures 3-5 It can be seen that this method is a significant improvement over existing technologies under different conditions, and also exhibits good robustness under different lighting conditions.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A polarization faint target enhancement detection method based on a bionic visual neural pathway, characterized in that, The method comprises the following steps: (1) acquiring a polarized image by using a polarized image sensor; (2) separating a static region from a dynamic region by using a frame difference method, and obtaining a dynamic region as a region where a weak dynamic target in the image is located; (3) calculating a degree of polarization of the dynamic region image by using a dense polarized solution method, and obtaining a degree of polarization image of the dynamic region; (4) using a bionic visual neural pathway to establish a visual enhancement model, the visual enhancement model comprising a photon receptor cell and a neuron, the photon receptor cell comprising an adaptive low-pass filter and an exponential low-pass filter in sequence, and the visual enhancement model being used for implementing image enhancement by performing a Naka-Rushton transformation on image features; weber The neuron is used as an input for performing a Naka-Rushton transformation on image features, so as to realize image enhancement; and the polarization degree image obtained in step (3) is input into the visual enhancement model, and an enhanced image is output. (5) performing connectivity analysis on the enhanced image to obtain the position of the weak target in the spatial domain.

2. The method of claim 1, wherein the method is based on a bionic visual neural pathway. In step (4), the adaptive low-pass filtering refers to dynamically adjusting the cut-off frequency of the low-pass filter and the variable gain control to change the filtering degree according to the signal strength.

3. The method of claim 2, wherein the method is based on a bionic visual neural pathway. In step (4), the specific operation of the adaptive low-pass filtering is as follows: where I L1 is the adaptive low-pass filtered image, f c1 denotes the angular frequency of this stage, f r denotes the frame rate, I DOP is the input polarimetric image; Normalization of the output of the low-pass filter, using the Naka-Rushton transformation to estimate the adaptive state I nr is represented as follows: I mid1 is the median of the intensity in the polarized image selected from the empirical dataset; The adaptive low-pass filtering is represented as follows: I L2 = f LPF (I DOP , f fm ) where f is the adaptive focusing frequency, calculated as: fm f = 1 / (T - T0) f fm = (f max -f min )I nr +f min f max and f min respectively represent the highest and lowest adaptation rates of the adaptive low-pass filter; Finally, a non-linear adaptive gain is used to compress the dynamic range, resulting in the image I after the adaptive low-pass filtering stage af The formula is as follows: I af = g af I L2 Gain factor g af Computed as: g af = (g max -1)(1-I nr )+1 where g max is the maximum gain factor.

4. The method of claim 2, wherein the method is based on a bionic visual neural pathway. In step (4), the exponential low-pass filtering process specifically comprises: The low-pass filtering loop is represented as follows: I L3 = f LPF (I af , f c3 ) where f c3 denotes the angular frequency of the phase, I L3 is the exponentially low-pass filtered output image; recompute the index low pass filtered result I weber : In the formula, alpha represents the sensitivity of the system.

5. The bionic visual neural pathway based polarized dim target enhancement detection method according to any one of claims 1 to 4, characterized in that, In step (4), the Naka-Rushton transformation adjusts the relationship between the output and the input to simulate the differential processing of signals of different intensities, and the adjustment formula is as follows: where I out represents the response value; I represents the intensity of the stimulating light; I mid2 represents the half-saturation stimulating intensity.

6. The bionic visual neural pathway based polarized dim target enhancement detection method according to any one of claims 1 to 4, characterized in that, In step (1), the polarized image sensor is a four-quadrant polarized camera.

7. The bionic visual neural pathway based polarized dim target enhancement detection method according to any one of claims 1 to 4, characterized in that, In step (2), a three-frame difference method is adopted to extract the dynamic region, so as to improve the proportion of the weak target in the image.

Images