A single-frame high dynamic range polarization imaging system and method

Abstract

The application discloses a single-frame high-dynamic polarization imaging system and method, wherein a linear polaroid is fixedly installed in front of a lens of a micro-polarization array (DoFP) camera, and a transmission axis of the linear polaroid is parallel to the direction of a micro-polarization array of a camera sensor chip. The scheme of the application is based on scene self-adaptation and channel decoupling, applied to high-dynamic polarization imaging in a complex light environment such as automatic driving target identification, and aims to solve the technical problem that polarization information is lost and target features are covered in a traditional micro-polarization array camera under an extremely strong light background, so that effective identification cannot be performed.

Description

Technical Field

[0001] This invention relates to the fields of computational optical imaging and machine vision technology, and in particular to a single-frame high dynamic polarization imaging. Background Technology

[0002] In the current fields of machine vision and autonomous driving, polarization imaging technology can reveal the physical properties (such as material and roughness) and three-dimensional geometric contour features of object surfaces that cannot be obtained by traditional two-dimensional light intensity images. It has irreplaceable advantages in dealing with severe weather and identifying highly reflective obstacles, and has therefore received widespread attention from academia and industry.

[0003] The Division of Focal Plane (DoFP) camera is the mainstream real-time polarization imaging device. Its basic physical mechanism involves directly integrating micro-polarizers with different transmission axis directions (typically 0°, 45°, 90°, and 135°) above the pixel array of the image sensor. Through this hardware architecture, the DoFP camera can simultaneously acquire light intensity information from four different polarization directions of the same scene in a single exposure. Subsequently, the system uses the light intensity data from these four channels to perform linear combination calculations to solve for the Stokes parameters and ultimately calculate the degree of polarization of each pixel in the image. However, in practical autonomous driving or outdoor industrial vision applications, the lighting conditions of the scene are often extremely complex. The environment often contains areas of extremely high brightness (such as oncoming headlights, strong specular reflections from smooth glass or water surfaces) and areas of extremely low illumination. This extreme "high dynamic range (HDR)" lighting environment poses a significant challenge to polarization calculation models that heavily rely on the absolute accuracy of the light intensity values ​​of each channel. This is a key background problem that polarization vision technology needs to overcome to move from ideal laboratories to complex industrial environments.

[0004] To acquire high dynamic range (HDR) images, some existing techniques utilize a standard camera to continuously capture multiple low dynamic range (LMR) images with different exposure levels, then derive the mapping relationship from sensor output to scene radiance for fusion synthesis. To obtain richer physical information, some existing techniques employ camera array schemes, such as using camera arrays equipped with different polarization filters, to synthesize HDR images by simultaneously acquiring multiple images. Furthermore, in recent years, innovative methods based on deep learning architectures have emerged, such as those utilizing residual dense networks, aiming to improve visual enhancement, object detection, and noise reduction and detail reconstruction of polarized images in low-light scenes.

[0005] Multi-frame sequential exposure fusion methods are cumbersome when processing multiple images, and their underlying photoelectric image sensors inherently have limited dynamic range in extreme high-light environments. When encountering severe overexposure due to extreme light (such as oncoming headlights, strong reflections from glass or metal surfaces), traditional polarization chip channels quickly reach full-well charge and saturate, leading to significant errors in the calculated polarization information and creating blind spots where information is completely lost in highly reflective areas of the image. While multi-camera array solutions can expand information acquisition channels to some extent, they significantly increase the hardware complexity and size of the system. Furthermore, although neural network methods have made significant progress in low-light and noise reduction, existing technologies still lack effective means to address the loss of underlying physical information caused by overexposure in extreme high-light conditions. This remains a significant obstacle to achieving reliable target recognition in complex high-light backgrounds.

[0006] In view of this, this invention is hereby proposed. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by proposing a single-frame high dynamic polarization imaging system and method. Based on scene adaptation and channel decoupling, it is applied to high dynamic polarization imaging in complex lighting environments such as autonomous driving target recognition. It aims to solve the technical problem that traditional micro-polarization array cameras lose polarization information and obscure target features due to extreme overexposure in extreme strong light backgrounds, thus failing to effectively identify targets.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A single-frame high dynamic range polarization imaging system includes a linear polarizer fixedly mounted in front of the lens of a micro-polarization array camera, with its transmission axis aligned with the micro-polarization array of the camera sensor chip. The direction is parallel.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A single-frame high dynamic range polarization imaging method, based on the system provided by this invention, includes the following steps: Step 1: Acquire a single-frame RAW image and separate the polarization channels to obtain sub-channel images L1, L2, L3, and L4; Step 2: Scene-adaptive effective extinction ratio Estimate; Step 3: Physical radiance restoration, obtaining light intensity data I1, I2, I3, I4; Step 4: Decouple the dual-model channels to obtain the high-exposure polarization degree. Low exposure polarization ; Step 5: HDR image fusion based on light intensity threshold.

[0010] Furthermore, step 1 includes the following steps: Step 1.1: Acquire a single frame of RAW image; Step 1.2: Based on the inherent spatial physical arrangement of 2×2 macropixels in the micro-polarization array, pixels with the same polarization direction in the original image are extracted at equal intervals to extract and separate pixels with a polarization angle of _____. , , , The sub-image channels L1, L2, L3, and L4.

[0011] Furthermore, step 2 includes the following steps: Step 2.1: Extract pixels with gray values ​​in the range [5, 245] from the sub-channel image and generate a safe pixel mask. ; Step 2.2: Calculate the real-time effective extinction ratio of the scene. The calculation formula is as follows: ; In the formula, mean() is the mean function, L1 is the sub-image channel with the highest transmittance because it is parallel to the transmission axis of the front polarizer, and L3 is the sub-image channel with the lowest transmittance because it is in an orthogonal extinction state.

[0012] Furthermore, in step 3, the effective extinction ratio obtained in step 2 is used... In addition, Malus's law is used to compensate for the grayscale image after channel separation, restoring the true physical incident light intensity of each channel. The formula is as follows: ; ; ; ; In the formula, , , , The polarization angles obtained from the separation are respectively: , , , Sub-image channels, This is the fixed compensation coefficient for the corresponding angle.

[0013] further, The value is 1 / 2.

[0014] Furthermore, in step 4, the calculation of the degree of polarization (DoP) is decoupled into two independent models for the performance of different channels under extreme lighting: the high-exposure model and the low-exposure model.

[0015] Furthermore, in step 4, the high-exposure model uses { for the highlighted areas. Light intensity data { Calculate the degree of polarization of high exposure The formula is as follows: ; ; ; ; In the formula, Total light intensity , For Stokes parameters.

[0016] Furthermore, in step 4, the low-exposure model targets dark areas using { Light intensity data Calculate the degree of polarization of low exposure The formula is as follows: ; ; ; ; In the formula, Total light intensity , For Stokes parameters.

[0017] Furthermore, in step 5, the brightest one... Using channel pixel values ​​as the criterion, pixel-level dynamic logical fusion is performed at each macro-pixel location: when When the grayscale value is measured, it is determined that the pixel is not in an extreme overexposure state, and a high-exposure polarization is used. When I1 > 240, it is determined that the pixel has encountered extremely strong reflection. The channel has lost its linear response; at this point, a smooth switch to low exposure polarization is necessary. The final output is a high dynamic polarization image of the entire scene.

[0018] Compared with the prior art, the beneficial effects of this invention are as follows: This invention aims to address the technical problem of traditional micro-polarization array cameras failing to effectively identify targets due to extreme overexposure in bright light environments. Specifically addressing the application requirement of target recognition in bright light, this embodiment prioritizes over obtaining absolutely accurate high-fidelity polarization values ​​under extreme glare. Instead, it focuses on intentionally attenuating incident light using extinction principles by adding a fixed polarizer in front of the lens, sacrificing some of the original scene's absolute polarization state. This significantly expands the system's dynamic range at the physical level. Through this design that transcends physical hardware and backend algorithms, this embodiment effectively recovers the target's structural outline within the extremely overexposed blind zone and extracts key geometric feature information, enabling the machine vision system to achieve highly robust target recognition even under various strong light interference backgrounds. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of a single-frame high dynamic range polarization imaging system according to Embodiment 1; Figure 2 This is a flowchart of a single-frame high dynamic range polarization imaging method according to Example 2; Figure 3 The original grayscale image was taken without a pre-polarizing filter. Figure 4 The original grayscale image captured by the system in Example 1; Figure 5 The DoP plot is calculated using the traditional method without a pre-polarizing filter; Figure 6 The Dop graph is calculated by the system in Example 1 using the method in Example 2; Figure 7 The image is a raw image taken without a pre-polarizing filter and contains severe highlights (such as reflections from a metal plate). Figure 8 The image is an original image captured by the system of Example 1, containing severe high-brightness reflections (such as reflections from a metal plate). Figure 9 This is the DoP image calculated using traditional methods in an overexposed scene without a pre-polarizing filter; Figure 10 The image shown is the DoP image calculated by the system in Example 1 using the method in Example 2 under an overexposed scene. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0021] Example 1: A single-frame high dynamic polarization imaging system, such as Figure 1 As shown in the schematic diagram of a single-frame high dynamic range polarization imaging system, a linear polarizer is fixedly mounted in front of the lens of the DoFP camera, and its transmission axis is aligned with the micro-polarization array of the camera sensor chip. The direction is parallel.

[0022] Definition: DoFP (Division of Focal Plane): A highly integrated polarization imaging chip architecture. In this chip, the surface of each pixel is covered with a miniature polarization filter, and the polarization filters covering four adjacent pixels are in different directions (typically 0°, 45°, 90°, and 135°), allowing the camera to simultaneously acquire polarization information of the scene in four different directions in a single snapshot exposure.

[0023] In this embodiment of the single-frame high dynamic range polarization imaging system, the transmission axis of the linear polarizer is aligned with the micro-polarization array of the camera sensor chip. When the direction is parallel, the incident light undergoes physical attenuation after passing through the pre-polarizer and is then captured by the camera sensor.

[0024] In this embodiment of the single-frame high dynamic range polarization imaging system, a linear polarizer is fixedly mounted in front of the lens of the DoFP camera, ensuring that the transmission axis of the linear polarizer is aligned with the micro-polarization array of the camera sensor chip. The parallel orientation creates an artificial light intensity gradient between channels, which is the basis for realizing the expansion of the physical level dynamic range. Specifically, it artificially creates a huge light intensity gradient difference between the four channels. This design is not a simple filter, but uses the physical extinction ratio to reduce the incident strong light in advance, so that even in extreme environments, the sensor that would otherwise be overexposed can retain at least one unsaturated channel data, providing a physical possibility for subsequent mathematical reconstruction.

[0025] Furthermore, in the single-frame high dynamic range polarization imaging system of this embodiment, the front-mounted fixed linear polarizer can be replaced with a polarizing prism (such as a Glan-Taylor prism), a beam-splitting polarizer, or a metal wire grid polarizer; alternatively, an electrically controlled liquid crystal phase delay unit or an electrically controlled liquid crystal tunable filter can be used to replace the static polarizer, dynamically adjusting the polarization direction or attenuation factor of the incident light by changing the applied voltage. These alternative methods share the common feature of achieving physical pre-attenuation and modulation of specific polarization components before the light enters the camera sensor, thereby protecting the sensor from instantaneous saturation by extremely strong light.

[0026] Example 2: A single-frame high dynamic polarization imaging method, such as Figure 2The flowchart shown (for a single-frame high dynamic polarization imaging method) includes the following steps: Step 1: Acquire a single-frame RAW image and separate the polarization channel to obtain sub-channel images L1, L2, L3, and L4.

[0027] Definition: RAW image: Raw digital image data captured by a machine vision sensor without undergoing internal image signal processing (ISP, such as white balance, gamma correction, etc.). Assuming it operates within the sensor's linear response region, it can be directly converted into normalized sensor irradiance domain data.

[0028] In this embodiment, step 1 includes the following steps: Step 1.1: Acquire a single frame RAW image.

[0029] In this embodiment, the single-frame RAW image obtained in step 1.1 is captured by the system in embodiment one.

[0030] Step 1.2: Based on the inherent spatial physical arrangement of 2×2 macropixels in the micro-polarization array, pixels with the same polarization direction in the original image are extracted at equal intervals to extract and separate pixels with a polarization angle of _____. , , , The sub-image channels L1, L2, L3, and L4.

[0031] In this embodiment, in step 1.2, four sub-image channels with different polarization angles are extracted and separated according to the pixel arrangement rules of the micro-polarization array.

[0032] Due to the physical effect of the pre-polarizer, the L1 channel has the highest transmittance (brightest) and the L3 channel has the lowest transmittance (darkest). Each adjacent 2×2 pixel unit constitutes a complete polarization detection unit, which is defined as a "macro-pixel" in this embodiment. It should be noted that since each macro-pixel corresponds to four pixels in the original image, the polarization degree (DoP) image generated subsequently is reduced to half of the original image in both the horizontal and vertical directions. That is, the resolution of the final output DoP image is 1 / 4 of the resolution of the original image.

[0033] Definitions: DoP (Degree of Polarization): A physical quantity describing the proportion of polarized light component to total light intensity in an incident light beam. Its calculated value ranges from [0, 1]. In this context, the precise DoP value in extreme glare is not the primary concern; rather, it serves as a key feature for extracting the target structure contours and information within the overexposure blind zone. Stokes Parameters: A set of mathematical parameters used to comprehensively describe the polarization state of a light beam from a macroscopic light intensity perspective. In this embodiment, it mainly involves... (Represents total light intensity) (Representing the difference in intensity between linearly polarized light at 0° and 90°) and (Represents the difference in intensity between linearly polarized light at 45° and 135°). These three parameters are the fundamental variables for solving the degree of polarization (DoP).

[0034] Step 2: Scene-adaptive effective extinction ratio Estimate.

[0035] Definition: ER (Extinction Ratio): One of the fundamental parameters characterizing the polarization performance of polarization devices (such as polarizers). It is defined as the maximum transmittance of linearly polarized light passing through a linear polarizer. ) and minimum transmittance ( The ratio of ) to ). In this scheme, it is used to assist in estimating the system's attenuation factor for incident light.

[0036] To adapt to complex and ever-changing real-world lighting environments, step 2 of this embodiment performs scene-adaptive effective extinction ratio. Estimation, used for real-time calibration.

[0037] In this embodiment, in step 2, the scene-adaptive effective extinction ratio is based on the "safe pixel mask". The real-time estimation algorithm abandons the traditional static calibration logic and dynamically calculates the current extinction ratio by automatically selecting "healthy pixels" in the image that are neither noise dead zones nor overexposed saturated areas. This real-time self-calibration process is the core of the solution's ability to adapt to dynamic lighting and solve dark divergence artifacts, ensuring that the algorithm can obtain accurate physical compensation parameters under different spectra and different incident angles.

[0038] In this embodiment, step 2 includes the following steps: Step 2.1: Extract pixels with gray values ​​in the range [5, 245] from the sub-channel image and generate a safe pixel mask. .

[0039] In this embodiment, step 2.1 can be replaced by a dynamic cropping method based on the statistical distribution of image histograms, for example, automatically removing pixels with gray values ​​at both ends of 5% based on the exposure distribution of each frame of the image.

[0040] Step 2.2: Calculate the real-time effective extinction ratio of the scene. The calculation formula is as follows: ; In the formula, mean() is the mean function, and L1 is the sub-image channel with the highest transmittance due to being parallel to the transmission axis of the pre-polarizer (i.e., the separated polarization angle is). L3 is the sub-image channel with the lowest transmittance due to being in an orthogonal extinction state (i.e., the separated polarization angle is...). (sub-image channels).

[0041] In this embodiment, step 2.1 can be replaced by a dynamic cropping method based on the statistical distribution of the image histogram, for example, automatically removing pixels with grayscale values ​​at the extremes of 5% based on the exposure distribution of each frame. Step 2.2 calculates the effective extinction ratio. The mean method can be replaced with the median method, mode method, or weighted average method to further enhance the robustness of the algorithm when facing complex texture scenes. The core purpose of these variations is the same: to select "healthy data" with high signal-to-noise ratio and linear response characteristics from the raw data for self-calibration.

[0042] Step 3: Physical radiance restoration, obtaining light intensity data I1, I2, I3, I4.

[0043] In this embodiment, in step 3, the effective extinction ratio obtained in step 2 is used as the basis for the calculation. In addition, Malus's law is used to compensate for the grayscale image after channel separation, restoring the true physical incident light intensity of each channel. The formula is as follows: ; ; ; ; In the formula, , , , The polarization angles obtained from the separation are respectively: , , , Sub-image channels, These are fixed compensation coefficients for the corresponding angles. Since the polarizer is parallel to the 0 channel in the experiment, we directly take 1 / 2 for these two compensation coefficients.

[0044] Glossary: ​​Malus's Law: A classic law of optical physics that describes the relationship between the intensity of transmitted light and the angle between the incident light's polarization plane and the polarizer's transmission axis after polarized light passes through a polarizer. In this scheme, the irradiance restoration step utilizes this law in conjunction with the extinction ratio to compensate for and restore the true incident light intensity of each polarization channel.

[0045] Step 4: Decouple the dual-model channels to obtain the high-exposure polarization degree. Low exposure polarization .

[0046] In this embodiment, step 4 constructs a "dual-model channel decoupling" calculation mechanism for different exposure levels. This breaks away from the traditional thinking that all four polarization channels must be used simultaneously for calculation. Through mathematical derivation, a decoupled model that can solve for the degree of polarization using only three channels is realized. In the high-exposure model, the darkest channel with extremely low signal-to-noise ratio is actively discarded, and in the low-exposure model, the brightest channel that has lost its linear response is actively discarded. This "refining" decoupling strategy completely avoids the computational divergence problem caused by extreme data.

[0047] In this embodiment, step 4, the "dual-model channel decoupling" calculation method breaks through the traditional thinking of full-channel joint calculation. It advocates selectively eliminating "failed channels" (such as saturated channels in the highlight area or noise channels in the shadow area) at different exposure levels, and only using the remaining effective polarization subset to solve for the Stokes vector and polarization degree. This channel redundancy utilization strategy for extreme lighting environments is the key core logic for solving computational divergence and eliminating information blind spots.

[0048] In this embodiment, in step 4, the calculation of polarization degree (DoP) is decoupled into two independent models for the performance of different channels under extreme lighting: the high-exposure model and the low-exposure model.

[0049] In this embodiment, in step 4, the high-exposure model targets bright areas and discards extremely dark areas that are easily drowned out by underlying noise. Channel, using { Light intensity data { Calculate the degree of polarization of high exposure The formula is as follows: ; ; ; ; In the formula, Total light intensity , For Stokes parameters.

[0050] In this embodiment, in step 4, the low-exposure model targets dark areas, discarding extremely bright areas that are most prone to charge trap overexposure. Channel, using { Light intensity data Calculate the degree of polarization of low exposure The formula is as follows: ; ; ; ; In the formula, Total light intensity , For Stokes parameters.

[0051] Step 5: HDR image fusion based on light intensity threshold.

[0052] In this embodiment, in step 5, a pixel-level HDR fusion strategy based on the brightest channel intensity threshold uses the restored physical light intensity I1 as the benchmark criterion to automatically and smoothly select the most suitable calculation model for the current illumination intensity at each pixel. This pixel-level dynamic switching mechanism allows the final synthesized image to retain the subtle structure of the dark areas while restoring the geometric contours in the bright blind areas, thereby maximizing the dynamic range of a single frame image at the algorithm level.

[0053] In this embodiment, step 5, the pixel-level HDR fusion strategy based on physical light intensity thresholds, encompasses a fusion process that uses the restored physical light intensity (especially the brightest I1 channel in the gradient sequence) as the criterion and achieves smooth switching between high and low exposure models at the pixel level. This single-frame processing method of "light intensity guidance and model switching" is a necessary technical means to ensure that the imaging results have both no motion artifacts and an ultra-wide dynamic range.

[0054] In this embodiment, in step 5, the brightest one is selected. Using channel pixel values ​​as the criterion, pixel-level dynamic logical fusion is performed at each macro-pixel location: when When the grayscale value (the threshold of 240 here is set according to the value range of 0-255 for an 8-bit quantized depth image) is determined to be not in an extreme overexposure state, a high-exposure polarization is adopted. When I1 > 240, it is determined that the pixel has encountered extremely strong reflection. The channel has lost its linear response; at this point, a smooth switch to low exposure polarization is necessary. The final output is a high dynamic polarization image of the entire scene.

[0055] In this embodiment, the "dual-model channel decoupling" mechanism in step 4 can be flexibly modified according to different sensor characteristics. For example, in addition to the specific three-channel subset mentioned in this embodiment, it can also be selected based on the actual noise level. or The Stokes parameters are solved by combining various methods. In the final image synthesis stage, the hard threshold pixel switching in step 5 can be replaced by a weighted smoothing fusion algorithm based on intensity mapping, making the transition between the high-exposure model and the low-exposure model smoother in the brightness transition region. These alternatives essentially utilize the mathematical redundancy between polarization channels to reliably extract the polarization features of the entire scene by avoiding overexposed saturation areas or low signal-to-noise ratio areas.

[0056] This embodiment aims to address the technical problem of traditional micro-polarization array cameras failing to effectively identify targets due to extreme overexposure in bright light environments. Specifically addressing the application requirement of target recognition in bright light, this embodiment prioritizes obtaining absolutely accurate high-fidelity polarization values ​​under extreme glare. Instead, it focuses on intentionally attenuating incident light using extinction principles by installing a fixed polarizer in front of the lens, sacrificing some of the original scene's absolute polarization state. This significantly expands the system's dynamic range at the physical level. Through this design that transcends physical hardware and backend algorithms, this embodiment effectively recovers the target's structural outline within the extremely overexposed blind zone and extracts key geometric feature information, enabling the machine vision system to achieve highly robust target recognition even under various strong light interference backgrounds.

[0057] The core idea of ​​this embodiment is to physically pre-attenuate the image by adding a fixed linear polarizer in a specific direction in front of the lens of a micro-polarization array camera, and combining scene-adaptive extinction ratio estimation with a dual-model channel decoupling algorithm. By strategically sacrificing absolute polarization fidelity, the limit of the sensor's physical dynamic range is broken, thus enabling the extraction of clear and robust target structure contours for machine vision systems even in extreme overexposure backgrounds such as strong light. This technical approach differs from traditional calibration methods that pursue high-precision physical measurements, emphasizing its practical invention purpose of achieving target perception and tracking by improving feature robustness in industrial vision fields such as autonomous driving target recognition and high-reflectivity background monitoring.

[0058] The technical advantage of this embodiment lies in its cross-dimensional combination of "physical pre-loss" and "mathematical reconstruction," which solves the information loss problem of traditional polarization cameras exhibiting "either black or bright" characteristics under extreme lighting conditions. By sacrificing a certain amount of original polarization fidelity at the hardware level, it successfully overcomes the physical limit of full-well charge in photoelectric sensors, enabling the system to acquire effective physical irradiance data even when encountering extreme glare such as direct headlights or strong reflections from metal or glass. This effectively eliminates the mathematical divergence caused by channel saturation in traditional polarization calculations, thus perfectly repairing the previously unsolvable black blind spots in the degree of polarization (DoP) image, ensuring the perceptibility and trackability of the target under extreme glare.

[0059] Furthermore, in terms of dynamic performance, the solution in this embodiment boasts the high efficiency of single-frame snapshot imaging, effectively avoiding motion artifacts. Compared to traditional solutions that synthesize high dynamic range images by repeatedly changing exposure times, this embodiment only requires a single exposure to acquire and fuse polarization information for the entire scene. This characteristic ensures that in scenarios with extremely high real-time requirements, such as autonomous driving and drone inspections, where targets are in high-speed motion, this solution will not produce image edge ghosting or "rainbow edges" or other artifacts caused by acquisition time differences, greatly improving the usability of images in dynamic environments.

[0060] Furthermore, the solution presented in this embodiment possesses excellent scene adaptability and robustness. Traditional polarized HDR solutions often rely on static calibration in a laboratory environment. Once the spectrum or angle of the incident light changes significantly, the fixed calibration coefficients become invalid, leading to severe noise divergence artifacts in dark areas. However, the solution presented in this embodiment, through its proposed "safe pixel mask" mechanism, can extract reliable data from each frame in real time and dynamically estimate the effective extinction ratio of the current scene, achieving real-time self-calibration of the algorithm. This means that the system can maintain stable output under various complex and variable lighting conditions (such as direct sunlight, alternating shadows, etc.) without repeatedly performing complex offline calibrations.

[0061] Finally, this embodiment provides extremely robust underlying feature support for backend machine vision algorithms. Under extremely harsh imaging conditions, this embodiment can not only recover the structural contours of target objects (such as window frame lines and vehicle shapes), but also significantly improve the problem of severe local distortion in polarization images. By recovering these originally lost key geometric information, this embodiment can significantly reduce the false detection rate and false negative rate of machine vision systems under high reflectivity and strong glare interference, providing solid technical support for cutting-edge tasks such as highly reflective road surface target recognition, cloud contour extraction under strong light backgrounds, and detection of highly reflective underwater objects.

[0062] Verification Example: To further help to intuitively understand the technical pain points of "overexposure" and "over-darkness" addressed by the solutions in Examples 1 and 2, the following comparative explanation is provided in conjunction with specific verification examples.

[0063] 1. Verification example for restoring details in areas with excessively dark shadows: refer to Figure 3 and 4 It is evident that in these two original images that were not processed by the algorithm of this scheme, due to the limitation of dynamic range, the characters on the plaques located in the shadow areas are extremely difficult to distinguish with the naked eye.

[0064] Figure 5 The DoP plot calculated using traditional methods without a pre-polarizer (i.e., the corresponding input) Figure 3 Information about the dark areas is still missing. Figure 6 The Dop graph (corresponding to the input) is calculated by the system of Example 1 using the method of Example 2. Figure 4 Through comparison, it is extremely obvious that... Figure 6 The limitations of dark noise were successfully overcome, and the words "Mathematics and" above "Interdisciplinary Science Research Institute" in the shadow were clearly presented, proving that the schemes in Examples 1 and 2 can still extract effective structured information in extremely dark areas.

[0065] 2. Verification example for contour recovery in overexposed areas under strong light: refer to Figure 7 and 8 In such extreme light conditions, traditional imaging channels are prone to local saturation.

[0066] Figure 9 The large pure black area (red box) in the image represents the mathematical divergence of Stokes' formula due to overexposure of local pixels (such as the 0° channel), resulting in the failure to solve for the degree of polarization (i.e., the formation of an "information blind spot / black hole"). Figure 10 The image shown is the DoP image synthesized using the physical attenuation and dual-model decoupling algorithm of this embodiment. It can be seen that the scheme in this embodiment perfectly eliminates the black blind spots originally caused by computational divergence, and successfully extracts and reconstructs the physical geometric contours of the metal plate even against a background of extremely strong high light reflection. This is of decisive significance for applications such as autonomous driving that require extremely high machine vision robustness.

[0067] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A single-frame high dynamic range polarization imaging system, characterized in that, A linear polarizer is fixedly mounted in front of the lens of a micro-polarization array camera, and its transmission axis is aligned with the micro-polarization array of the camera sensor chip. The direction is parallel.

2. A single-frame high dynamic range polarization imaging method, characterized in that, The system based on claim 1 includes the following steps: Step 1: Acquire a single-frame RAW image and separate the polarization channels to obtain sub-channel images L1, L2, L3, and L4; Step 2: Scene-Adaptive Effective Extinction Ratio Estimate; Step 3: Physical radiance restoration, obtaining light intensity data I1, I2, I3, I4; Step 4: Decouple the dual-model channels to obtain the high-exposure polarization degree. Low exposure polarization ; Step 5: HDR image fusion based on light intensity threshold.

3. The single-frame high dynamic range polarization imaging method according to claim 2, characterized in that, Step 1 includes the following steps: Step 1.1: Acquire a single frame of RAW image; Step 1.2: Based on the inherent spatial physical arrangement of 2×2 macropixels in the micro-polarization array, pixels with the same polarization direction in the original image are extracted at equal intervals to extract and separate pixels with a polarization angle of _____. , , , The sub-image channels L1, L2, L3, and L4.

4. The single-frame high dynamic range polarization imaging method according to claim 2, characterized in that, Step 2 includes the following steps: Step 2.1: Extract pixels with gray values ​​in the range [5, 245] from the sub-channel image and generate a safe pixel mask. ; Step 2.2: Calculate the real-time effective extinction ratio of the scene. The calculation formula is as follows: ； In the formula, mean() is the mean function, L1 is the sub-image channel with the highest transmittance because it is parallel to the transmission axis of the front polarizer, and L3 is the sub-image channel with the lowest transmittance because it is in an orthogonal extinction state.

5. The single-frame high dynamic range polarization imaging method according to claim 1, characterized in that, In step 3, the effective extinction ratio obtained in step 2 is used as a basis. In addition, Malus's law is used to compensate for the grayscale image after channel separation, restoring the true physical incident light intensity of each channel. The formula is as follows: ； ； ； ； In the formula, , , , The polarization angles obtained from the separation are respectively: , , , Sub-image channels, This is the fixed compensation coefficient for the corresponding angle.

6. The single-frame high dynamic range polarization imaging method according to claim 5, characterized in that, The value is 1 / 2.

7. The single-frame high dynamic range polarization imaging method according to claim 1, characterized in that, In step 4, the calculation of polarization degree is decoupled into two independent models to address the performance of different channels under extreme lighting conditions: the high-exposure model and the low-exposure model.

8. A single-frame high dynamic range polarization imaging method according to claim 7, characterized in that, In step 4, the high-exposure model uses { for the highlighted areas. Light intensity data { Calculate the degree of polarization of high exposure The formula is as follows: ； ； ； ； In the formula, Total light intensity , For Stokes parameters.

9. A single-frame high dynamic range polarization imaging method according to claim 7, characterized in that, In step 4, the low-exposure model targets dark areas using { Light intensity data Calculate the degree of polarization of low exposure The formula is as follows: ； ； ； ； In the formula, Total light intensity , For Stokes parameters.

10. A single-frame high dynamic range polarization imaging method according to claim 1, characterized in that, In step 5, the brightest one... Using channel pixel values ​​as the criterion, pixel-level dynamic logical fusion is performed at each macro-pixel location: when At that time, it was determined that the pixel was not in an extreme overexposure state, and a high-exposure polarization was adopted. When I1 > 240, it is determined that the pixel has encountered extremely strong reflection. The channel has lost its linear response; at this point, a smooth switch to low exposure polarization is necessary. The final output is a high dynamic polarization image of the entire scene.

Images