A real-time image stabilization method and system for intelligent image processing

Abstract

The application provides a real-time image stabilization method and system for intelligent image processing, and relates to the technical field of image stabilization. The application synchronously acquires continuous image frames under strong glare interference, and identifies a texture blur area. A texture complexity distribution map is calculated for the area in combination with a local structure change amount of adjacent frames. A candidate area is screened according to the texture complexity distribution map. Polarization filtering processing is applied to the candidate area. A polarization filtering angle is dynamically adjusted to match the polarization direction of incident light, and a polarization filtering image is generated. A motion vector field of the candidate area is extracted based on the gradient continuity feature of the image. The motion vector field is spatially superimposed with the texture complexity distribution map to form a motion vector set carrying a texture weight factor. Background vector groups and foreground abnormal groups are separated through hierarchical clustering according to the texture weight factor. Only the geometric center value of the background groups is selected to perform inter-frame motion compensation, and a stable image sequence is output. The application improves the stabilization quality of images collected by an unmanned intelligent turntable in a strong glare dynamic scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image stabilization, in particular to a real-time image stabilization method and system for intelligent image processing. BACKGROUND

[0002] In a complex light environment, the imaging device mounted on an unmanned intelligent turntable often produces high-frequency image jitter due to severe pitching and rotating motion and environmental strong glare interference, resulting in inaccurate motion target tracking and distorted scene recognition. Such a scene urgently needs a real-time image stabilization technology that can overcome the texture feature degradation caused by dynamic glare and the inter-frame displacement distortion caused by turntable motion, while maintaining low processing delay to meet real-time control requirements.

[0003] The current mainstream solution uses multi-scale optical flow field fusion and adaptive motion filtering technology: by extracting the dense optical flow field of adjacent frames, the motion vector level is divided in combination with scene depth information; then a space-time filter is constructed based on the statistical distribution of motion vector amplitude to suppress high-frequency abnormal motion components; finally, inter-frame compensation is achieved through global motion model fitting. This solution relies on GPU acceleration to ensure the real-time performance of optical flow calculation.

[0004] This solution has double failures in strong glare environment: glare overwhelms local texture structure, resulting in a large number of erroneous estimation points in the texture sparse area of the optical flow field, and the motion vector level division is inaccurate; the transient light intensity change caused by environmental glare is misidentified as foreground motion noise, the space-time filter incorrectly filters out the real turntable motion component, causing distortion of the compensation model. This defect is particularly pronounced in high-reflectivity scenes, and the stabilized image still has periodic jitter and edge tearing. SUMMARY

[0005] The present application provides a real-time image stabilization method and system for intelligent image processing to solve the problem of low stabilization quality of images collected by an unmanned intelligent turntable in the prior art.

[0006] In a first aspect, the present application provides a real-time image stabilization method for intelligent image processing, comprising:

[0007] When the unmanned intelligent turntable performs pitching or rotating motion, continuously acquiring image frames disturbed by strong environmental glare, and identifying texture blur areas caused by light reflection in the continuous image frames;

[0008] For the texture blur area, calculate a texture complexity distribution map in combination with the local structure change amount of adjacent image frames, and select a candidate area in the texture blur area with a texture feature retention rate higher than a preset standard based on the texture complexity distribution map;

[0009] applying a polarized light filtering process on the candidate region, generating a polarized light filtering image with gradient continuity feature of suppressing glare noise and enhancing edge gradient by adjusting filtering angle to match incident light polarization direction of current frame;

[0010] extracting a motion vector field of the candidate region between adjacent frames based on the gradient continuity feature of the polarized light filtering image, and spatially superimposing the motion vector field and the texture complexity distribution map to form a motion vector set carrying a texture weight factor;

[0011] performing hierarchical clustering on the motion vector set according to the texture weight factor, separating a background motion vector group generated by the motion of the unmanned intelligent turntable itself and a foreground abnormal vector group caused by environmental glare interference, and only selecting the geometric center value of the background motion vector group for inter-frame motion compensation to output a stable image sequence.

[0012] Optionally, the extracting a motion vector field of the candidate region between adjacent frames based on the gradient continuity feature of the polarized light filtering image, and spatially superimposing the motion vector field and the texture complexity distribution map to form a motion vector set carrying a texture weight factor comprises:

[0013] tracking the displacement trajectory of the corresponding edge points between adjacent frames along the object edge direction in the candidate region of the polarized light filtering image, and constructing a motion vector field through the direction and length of the displacement trajectory;

[0014] binding the spatial position of each motion vector in the motion vector field with the value of the texture complexity distribution map at the same spatial position;

[0015] assigning a texture weight factor to each motion vector according to the bound value, the texture weight factor being in a positive proportional relationship with the value of the texture complexity distribution map, to form a motion vector set carrying a texture weight factor.

[0016] Optionally, the applying a polarized light filtering process on the candidate region, generating a polarized light filtering image with gradient continuity feature of suppressing glare noise and enhancing edge gradient by adjusting filtering angle to match incident light polarization direction of current frame comprises:

[0017] analyzing the dominant polarization direction of incident light in the candidate region according to the light reflection characteristics of the candidate region, the dominant polarization direction being determined by the periodic change pattern of pixel brightness in consecutive frames;

[0018] dynamically rotating the light transmission axis angle of the physical filter to form an orthogonal relationship between the light transmission axis angle and the dominant polarization direction;

[0019] According to the adjusted polarizing axis angle, light rays passing through the candidate region of the physical filter are collected to generate a polarized filter image, in which the luminance amplitude of glare noise is reduced and the continuity of light and dark transition of object edges is improved.

[0020] Optionally, the motion vector set is hierarchically clustered according to the texture weight factor to separate a background motion vector group caused by the motion of the unmanned intelligent turntable itself and a foreground abnormal vector group caused by environmental glare interference, and only the geometric center value of the background motion vector group is selected for inter-frame motion compensation to output a stable image sequence, including:

[0021] The motion vector set is divided into a high weight group and a low weight group according to the texture weight factor, the high weight group contains motion vectors with a texture weight factor higher than a preset separation threshold, and the low weight group contains the remaining motion vectors;

[0022] The direction consistency density of the motion vectors in the high weight group is calculated, the motion vectors falling into the density peak value region are classified as a background motion vector group, and the motion vectors in the low weight group and the high weight group not falling into the density peak value region are classified as a foreground abnormal vector group;

[0023] Only the geometric center value of all vectors in the background motion vector group is selected as a translation reference to perform reverse displacement compensation on the next frame of image to output a stable image sequence.

[0024] Optionally, for the texture blur region, a texture complexity distribution map is calculated in combination with the local structure change amount of the adjacent image frames, and a candidate region with a texture feature retention rate higher than a preset standard in the texture blur region is selected based on the texture complexity distribution map, including:

[0025] For each image cell in the texture blur region, the structure change amount of the corresponding image cells between the current frame and the adjacent frame is calculated, and the structure change amount is represented by the cumulative value of the difference amplitude of the pixel values within the cell;

[0026] The structure change amount is mapped to each image cell position of the texture blur region to generate a texture complexity distribution map, and the value of the texture complexity distribution map reflects the strength of the ability of the image cell to resist glare interference;

[0027] In the texture complexity distribution map, image cells with a value higher than a preset threshold are marked, and continuously distributed marked cells in the image cells are merged into a candidate region, and the candidate region covers a continuous spatial range with a texture feature retention rate higher than a preset standard.

[0028] Optionally, the dynamic rotating physical filter's light transmission axis angle is adjusted to form an orthogonal relationship with the dominant polarization direction, including:

[0029] An angle value of the dominant polarization direction is measured, and an orthogonal angle gap between the angle value and the light transmission axis angle is calculated;

[0030] A motor driving pulse number is generated according to the orthogonal angle gap, and a rotating motor is driven to rotate the physical filter by a gap compensation angle;

[0031] During the rotation, when the light transmission axis angle sensor feedback value reaches the orthogonal angle of the angle value, the rotating motor is stopped and the light transmission axis is locked.

[0032] Optionally, the bound values are used to assign texture weight factors to each motion vector, the texture weight factors are in a proportional relationship with the values of the texture complexity distribution map, and a motion vector set carrying texture weight factors is formed, including:

[0033] The bound values and the texture complexity distribution map values corresponding to the motion vector spatial positions are used as texture reliability values of the current motion vector;

[0034] The texture reliability values are input into a linear converter to generate texture weight factors through fixed amplification operation, and the values of the texture weight factors are equal to the texture reliability values multiplied by a preset proportion coefficient;

[0035] The texture weight factors are written into the attribute field of the corresponding motion vector to construct the motion vector set carrying the texture weight factors.

[0036] In a second aspect, the present application provides a real-time image stabilization system for intelligent image processing, including:

[0037] An acquisition module is configured to acquire continuous image frames disturbed by environmental strong glare light when an unmanned intelligent turntable performs pitching or rotating motion, and identify texture blur areas caused by light reflection in the continuous image frames;

[0038] A calculation module is configured to calculate a texture complexity distribution map for the texture blur areas in combination with local structure change amounts of adjacent image frames, and filter candidate areas with a texture feature retention rate higher than a preset standard in the texture blur areas based on the texture complexity distribution map;

[0039] A processing module is configured to apply polarization filtering processing to the candidate areas, generate a polarization filtering image by adjusting a filtering angle to match an incident light polarization direction of a current frame, and the polarization filtering image has both glare noise suppression and gradient continuity enhancement of edge gradient features;

[0040] extracting a motion vector field of the candidate region between adjacent frames based on gradient continuity features of the polarized light filtering image, and spatially superimposing the motion vector field and the texture complexity distribution map to form a motion vector set carrying a texture weight factor;

[0041] a separation module configured to hierarchically cluster the motion vector set according to the texture weight factor, separate a background motion vector group generated by the motion of the unmanned intelligent turntable and a foreground abnormal vector group caused by environmental glare interference, and only select a geometric center value of the background motion vector group for inter-frame motion compensation to output a stable image sequence.

[0042] In a third aspect, the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the real-time image stabilization method for intelligent image processing according to the first aspect.

[0043] In a fourth aspect, the present application provides a computer storage medium storing a computer program, which, when executed by a computer, implements the real-time image stabilization method for intelligent image processing according to the first aspect.

[0044] In the embodiments of the present application, when the unmanned intelligent turntable performs pitching or rotating motion, continuous image frames disturbed by environmental strong glare are synchronously acquired, and texture blurred regions caused by light reflection in the continuous image frames are identified; for the texture blurred regions, a texture complexity distribution map is calculated in combination with a local structure change amount of adjacent image frames, and a candidate region with a texture feature retention rate higher than a preset standard is selected based on the texture complexity distribution map; a polarized light filtering process is applied to the candidate region, a polarized light filtering image is generated by adjusting a filtering angle to match an incident light polarization direction of a current frame, the polarized light filtering image has gradient continuity features of suppressing glare noise and enhancing edge gradient; a motion vector field of the candidate region between adjacent frames is extracted based on the gradient continuity features of the polarized light filtering image, and the motion vector field is spatially superimposed with the texture complexity distribution map to form a motion vector set carrying a texture weight factor; the motion vector set is hierarchically clustered according to the texture weight factor, a background motion vector group generated by the motion of the unmanned intelligent turntable and a foreground abnormal vector group caused by environmental glare interference are separated, and only a geometric center value of the background motion vector group is selected for inter-frame motion compensation to output a stable image sequence.

[0045] The technical scheme of the present application has the following beneficial effects:

[0046] The texture degradation area in the image under the interference of the precisely positioned strong glare is located, a target is provided for subsequent processing, and waste of global computing resources is avoided. Through texture complexity quantification and screening, candidate areas with strong anti-glare ability are extracted, and the reliability of the motion estimation basis feature is ensured. Dynamic polarization filtering generates an optimized image for glare suppression and edge enhancement, directly solving the texture blur problem and providing high continuity features for motion tracking. The texture reliability and motion vector are fused to form a weighted motion set, and an anti-interference motion analysis input is constructed. Based on the texture weight clustering, the precise separation of the real motion of the turntable and the glare pseudo-motion is realized, and the physical consistency of the motion compensation is ensured.

[0047] Further, in the candidate area of the polarization filtering image, the motion vector field is constructed by tracking the displacement trajectory of the adjacent frame edge points along the object edge; the spatial position of each motion vector is bound with the texture complexity distribution map value at the same position; the motion vector is given a texture weight factor in direct proportion to the bound value, and a weighted motion vector set is formed. Through the spatial binding and weight mapping of the motion vector and the texture complexity, the texture anti-interference ability is converted into the motion vector reliability index, which provides a decision basis for distinguishing real motion from glare pseudo-motion for subsequent clustering.

[0048] These aspects or other aspects of the present application will be more apparent in the following description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0049] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0050] Figure 1 A flow chart of a real-time image stabilization method for intelligent image processing provided by the present application is shown;

[0051] Figure 2 A flow chart of a real-time image stabilization method for intelligent image processing provided by the present application is shown;

[0052] Figure 3 A structural schematic diagram of a real-time image stabilization system for intelligent image processing provided by the present application is shown;

[0053] Figure 4 A structural schematic diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0054] In order for those skilled in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0055] In some of the processes described in the specification and claims of the present application and the above-described drawings, a plurality of operations are included which occur in a specific order, but it should be clearly understood that these operations can be performed in an order other than that in which they appear or in parallel, and the serial numbers of the operations, such as 101, 102, etc., are only used to distinguish different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes can include more or fewer operations, and the operations can be performed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this paper are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do "first" and "second" represent different types.

[0056] The current image stabilization scheme based on multi-scale optical flow field fusion and adaptive motion filtering has fundamental limitations when the unmanned intelligent turntable encounters a strong glare scene: first, the glare overwhelms the local texture structure, causing the optical flow field to generate a large number of false motion estimation points in the texture sparse area, destroying the physical reality of the motion vector level division; second, the transient light intensity changes caused by environmental glare are misidentified as foreground motion noise, causing the space-time filter to incorrectly filter out the real motion component of the turntable. This double defect causes the compensation model to output distorted images with periodic jitter and edge tearing in high-reflectivity scenes, severely restricting the reliability of real-time operations in complex lighting environments.

[0057] In view of the above defects, the present application proposes a real-time stabilization method based on texture-polarization-motion cooperation, and the core innovation is that: through the cross-modal coupling of dynamic polarization filtering and texture anti-interference capability, the reliability basis of motion estimation is reconstructed. Specifically, under strong glare interference, first, select candidate areas with high texture feature retention rate and apply polarization angle adaptive filtering processing to generate an optimized image with continuous edge gradients; then convert the texture complexity into a motion vector weight factor to drive the hierarchical clustering to accurately separate the real motion of the turntable from the glare false motion; finally, perform inter-frame compensation based on the geometric center value of the background motion group. This method has broken through the problem of motion estimation distortion caused by texture feature degradation in glare scenes - both eliminating false estimates of the optical flow field in texture sparse areas and avoiding false filtering of motion components caused by transient light changes, fundamentally overcoming the periodic jitter and edge tearing defects in existing technologies, while meeting the millisecond-level real-time processing requirements.

[0058] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0059] Figure 1 A flowchart of a real-time image stabilization method for intelligent image processing is provided in the embodiments of the present application, as shown in Figure 1 The method comprises the following steps.

[0060] 101. When an unmanned intelligent rotating platform performs pitching or rotating motion, continuously acquire image frames disturbed by strong environmental glare, and identify texture blur areas caused by light reflection in the image frames.

[0061] In the above scheme, the unmanned intelligent rotating platform is a rotatable platform carrying a camera device, which can perform pitching or rotating motion. The strong environmental glare disturbance is a dazzling light spot formed by strong natural light or artificial light on the surface of a high-reflectivity object such as metal or water. The continuous image frames are a plurality of images continuously captured in time sequence during the motion of the rotating platform. The texture blur area is an area in which the details such as the surface texture and edge of an object are lost due to glare coverage in the image.

[0062] In the embodiments of the present application, first, the image acquisition module is triggered by the attitude sensor built in the rotating platform, and a high-speed camera is started at the moment when the pitching or rotating motion of the rotating platform starts, to capture continuous images at a rate of 30 frames per second. If there is a strong glare source in the environment at this time, the camera sensor will capture an image sequence containing overexposed spots.

[0063] Secondly, the acquired continuous image frames are analyzed by region scanning: each image is divided into a plurality of square units, such as 16x16 pixel blocks; the pixel value fluctuation amplitude in each unit is calculated, that is, the difference between the brightest and darkest pixels in the unit; the unit with a fluctuation amplitude lower than a set threshold is marked as a low-texture unit, and such area loses the contrast between light and dark due to glare coverage, presenting a uniform bright spot; the low-texture units that are continuously adjacent in space are merged into a texture blur area. Finally, the boundary coordinate set of all the texture blur areas is output as the targeted target for subsequent processing.

[0064] 102. For the texture blur area, calculate a texture complexity distribution map in combination with the local structure change amount of the adjacent image frames, and filter a candidate area with a texture feature retention rate higher than a preset standard from the texture blur area based on the texture complexity distribution map.

[0065] Optionally, as shown in Figure 2 Step 102 can specifically include the following steps:

[0066] 1021. For each image unit within the texture blurring region, calculate the structural change of the corresponding image unit between the current frame and adjacent frames. The structural change is represented by the cumulative value of the difference in pixel values ​​within the unit.

[0067] 1022. Map the structural change amount to the position of each image unit in the texture blur region to generate a texture complexity distribution map. The value of the texture complexity distribution map reflects the strength of the image unit's ability to resist glare interference.

[0068] 1023. In the texture complexity distribution map, the image units marked with values ​​higher than a preset threshold are merged into candidate regions, and the candidate regions cover a continuous spatial range with a texture feature retention rate higher than a preset standard.

[0069] In the above scheme, an image unit refers to a fixed-size square region into which a textured, blurred area is divided. The structural change refers to the sum of the brightness changes of all pixels in the same image unit across two adjacent frames; a larger value indicates more significant changes in texture detail. The texture complexity distribution map is a mapping map that labels the structural change value of each image unit. High-value regions represent locations with strong anti-glare capabilities. Candidate regions are continuous blocks formed by connecting adjacent high-value image units; these regions can still retain the true texture features of the object under glare interference.

[0070] In this embodiment, step 1021 first divides the textured blurry region identified in step 101 into squares of fixed size. For each square, the following operations are performed: Locating corresponding squares in adjacent frames: Based on the turntable motion parameters and pitch / rotation angles, aligning the current frame with the same physical region of the previous frame; Calculating pixel brightness differences: Comparing the brightness values ​​of corresponding positions within the squares of the two frames pixel by pixel, and calculating the absolute difference. For example, if the brightness of pixel A changes from 120 to 115, the difference = |120-115| = 5; Accumulating the difference values: Adding the brightness differences of all pixels within the square to obtain the structural change. The larger the total brightness change, the richer the texture details in the region and the less affected by glare.

[0071] Then, spatial mapping is performed in step 1022: coordinate labels are generated for each square, and the value of its structural change is marked at the corresponding coordinate position; a distribution map is generated: the values ​​of all squares are arranged according to their spatial positions to form a two-dimensional numerical matrix; capability strength correlation: areas with values ​​≥1000 in the distribution map represent strong anti-glare capability and clear texture, while areas with values ​​≤500 represent weak anti-glare capability and glare that overwhelms the texture.

[0072] Finally, step 1023 is performed for threshold marking: set the anti-glare ability threshold, mark all squares with values ≥1000 in the distribution map; adjacent merging: check the continuity of adjacent marked squares, i.e. up, down, left and right connections, and merge the continuous blocks into candidate regions, for example, three adjacent squares with values 1280, 1200 and 1100 are merged into an L-shaped region); and region output: the candidate region needs to meet the minimum area constraint to ensure covering the effective texture area.

[0073] In actual application, the vehicle travels on a mountain road at night at 60 km / h, and the vehicle-mounted camera continuously shakes due to the uneven road surface. The front motorcycle taillight forms a candidate region in the monitoring picture. The region is divided into 256 16×16 pixel squares through step 1021. For a square at the left edge of the bumper: the average pixel brightness in the current frame is 170, and that in the previous frame is 163; the maximum pixel brightness difference is |182-175|=7, and the minimum difference is |155-152|=3; the total brightness difference of 256 pixels is 1350, which is high due to the change of reflected light caused by the concave indentation of the bumper.

[0074] Step 1022 marks 1350 on the coordinates (120, 80) of the square in the scene, and the adjacent glare center square, which is located at the middle of the bumper, has a total brightness difference of only 280, and the pixel brightness under the glare is almost unchanged. The distribution map shows that the square at the edge of the bumper has a value of 1100-1500, indicating high anti-glare ability; and the square at the center of the glare has a value of 200-400, indicating low anti-glare ability.

[0075] Step 1023 sets the threshold value to 1000, and marks the square with high value. Twelve squares with values 1100-1500, which are continuously distributed at the edge of the bumper, are merged to form a U-shaped candidate region surrounding the glare spot, which clearly retains the license plate characters and screw hole texture. In the case of continuous vehicle shaking, the same physical region is shifted by 5 pixels in adjacent frames due to vibration. After correcting the coordinates through the turntable motion parameters, the structure change calculation is not affected by the displacement.

[0076] The overall scheme of the above step 102, when the high-speed moving carrier is subjected to the dual interference of vibration and glare, accurately locks the physical region least affected by glare by quantifying the dynamic change characteristics of the local texture, builds an anti-interference anchor point basis for subsequent motion analysis, and fundamentally avoids the feature distortion problem caused by vibration displacement and strong light flooding.

[0077] 103. Apply polarization filtering to the candidate region, adjust the filtering angle to match the incident light polarization direction of the current frame, generate a polarization filtered image, and the polarization filtered image has the characteristics of suppressing glare noise and enhancing edge gradient gradient continuity;

[0078] Optionally, step 103 can specifically include the following steps:

[0079] 1031. Analyzing a dominant polarization direction of the incident light in the candidate region according to the light reflection characteristics of the candidate region, the dominant polarization direction being determined by a periodic change pattern of pixel brightness in consecutive frames;

[0080] 1032. Dynamically rotating an angle of a light transmission axis of the physical filter to form an orthogonal relationship between the angle of the light transmission axis and the dominant polarization direction;

[0081] In some embodiments, the step 1032 can include the following processes: measuring an angle value of the dominant polarization direction, calculating an orthogonal angle gap between the angle value and the angle of the light transmission axis; generating a number of motor driving pulses according to the orthogonal angle gap, driving a rotating motor to rotate the physical filter by a gap compensation angle; and stopping the rotating motor and locking the light transmission axis when a feedback value of the light transmission axis angle sensor reaches the orthogonal angle of the angle value during the rotation.

[0082] 1033. Collecting light rays passing through the candidate region of the physical filter according to the adjusted angle of the light transmission axis to generate a polarized light filtering image, in which a brightness amplitude of glare noise is reduced and a light-dark transition continuity of object edges is improved.

[0083] In the above solutions, the dominant polarization direction refers to a main vibration angle of reflected light in the candidate region. The angle of the light transmission axis refers to a direction angle of the physical filter allowing light to pass through. The orthogonal relationship refers to the light transmission axis being perpendicular to the polarization direction. The brightness amplitude refers to a range of sharp fluctuations of pixel brightness in the glare region. The light-dark transition continuity refers to a smoothness of color depth changes at the edges of the object.

[0084] In some embodiments, first, the step 1031 extracts a brightness sequence of consecutive frames, selects brightness values of each pixel in the candidate region in consecutive 5 frames of images to generate a time-dimension brightness sequence, calculates a brightness fluctuation direction: calculates a sign change rule of adjacent frame brightness difference values for each pixel: if the brightness value presents a periodic fluctuation of "low-high-low-high", records the fluctuation direction of the pixel as 0°; if it presents a fluctuation of "high-low-high-low", records it as 90°; and other rules are regarded as no direction. Statistics of the dominant direction: classifies all effective directions in the candidate region into a 0°-180° angle interval, and counts a median angle of the highest proportion group as the dominant polarization direction. For example, a 5-frame brightness of a certain pixel [50, 120, 55, 125, 60]→difference sequence [+70, -65, +70, -65]→sign sequence [+,-,+,-]→fluctuation direction 0°.

[0085] Subsequently, the target angle is calculated by step 1032: when the dominant polarization direction is a, the target angle of the light transmission axis = a + 90°, and if it exceeds 180°, it is reduced by 180°. The rotation mechanism is driven: the current angle of the filter is read β; the shortest rotation path is calculated: if |a + 90° - β| > 180°, rotate in the opposite direction; send pulse instructions: each pulse rotates 0.1°, and the total number of pulses = angle difference / 0.1. Real-time closed-loop calibration: the actual angle is fed back in real time through the angle sensor, and the rotation stops when the error with the target angle is <1°. For example, the dominant direction a = 40° → target angle = 130°; the current angle β = 20° → 110° needs to be turned → 1100 pulses are sent.

[0086] Finally, the light is physically filtered by step 1033: when the light transmission axis is orthogonal to the polarization direction, the main component of glare is blocked. Weak light compensation exposure: because the filter reduces the amount of incoming light, the exposure time is automatically extended by 1.5 times. Image generation verification: glare area: the brightness value changes from severe fluctuations to stability; object edge: the original broken gray jump becomes a continuous transition.

[0087] In practical applications, assume that a monitoring camera captures a wet and slippery road surface at night, and the car lights form a candidate area at the water accumulation. By analyzing 5 consecutive frames with a 0.05 second interval through step 1031: the pixel A brightness sequence: [180, 50, 190, 55, 200] → difference [+130, -140, +135, -145] → sign [+,-,+,-] → fluctuation direction 0°; 68% of the pixels in the area fluctuate in the 0° direction, and the dominant polarization direction = 0°.

[0088] Through step 1032, the target angle of the light transmission axis = 0° + 90° = 90°; the current angle sensor reading is 170°, the shortest path: counterclockwise 80°, 170° → 180° → 0° → 90°; 800 pulses are sent, i.e. 80° / 0.1°, and the offset is 3° due to vibration during rotation, and 30 pulses are added for compensation → the final angle is 90.2°.

[0089] After filtering by step 1033, the image is collected: the brightness of the water reflection decreases from an average of 200 to 60; the original zebra crossing edge, which was submerged by the glare of the car lights, appears, and the gray value changes from [25, 155] to [65, 68, 71]; the details of the pedestrian shoe texture appear, and the original brightness difference <5, now >15. Weak light compensation: the exposure time is extended from 1 / 100 seconds to 1 / 67 seconds, and the dark part noise does not increase significantly.

[0090] The overall scheme of the above steps 103, in the night scene where weak light and strong glare coexist, accurately suppresses the interference reflection of artificial light sources through dynamic polarization filtering, synchronously enhances the visual continuity of object edges in low-illumination environments, provides clear image foundation for motion target tracking that is not affected by environmental light pollution, and significantly improves the stability and reliability of night monitoring systems.

[0091] 104. based on the gradient continuity feature of the polarization filtered image, extracting a motion vector field of the candidate region between adjacent frames, and spatially superimposing the motion vector field and the texture complexity distribution map to form a motion vector set carrying a texture weight factor;

[0092] Optionally, step 104 can specifically include the following steps:

[0093] 1041. in the candidate region of the polarization filtered image, tracking the displacement trajectory of the corresponding edge points between adjacent frames along the object edge direction, and constructing a motion vector field through the direction and length of the displacement trajectory;

[0094] 1042. binding the spatial position of each motion vector in the motion vector field with the value of the texture complexity distribution map at the same spatial position;

[0095] 1043. according to the bound value, assigning a texture weight factor to each motion vector, the texture weight factor being in a positive proportional relationship with the value of the texture complexity distribution map, to form a motion vector set carrying a texture weight factor.

[0096] In step 1043, the bound value and the texture complexity distribution map value corresponding to the spatial position of the motion vector can be used as the texture reliability value of the current motion vector; the texture reliability value can be input into a linear converter to generate a texture weight factor through a fixed amplification operation, the value of the texture weight factor being equal to the texture reliability value multiplied by a preset proportionality coefficient; the texture weight factor can be written into the attribute field of the corresponding motion vector to construct a motion vector set carrying a texture weight factor.

[0097] In the above scheme, the gradient continuity feature is the smooth transition property of the color depth change of the object edge. The displacement trajectory is the moving path of the same object feature point in two frames of pictures. The motion vector field is a set of multiple directional arrows, each arrow representing the moving direction and distance of a feature point. Spatial binding is to establish a corresponding relationship between the position coordinates of the arrow and the value of the texture anti-interference ability at the same position. The texture weight factor is a value that measures the reliability of the motion vector, and the stronger the texture anti-interference ability, the higher the value.

[0098] In the embodiments of the present application, first, the edge point positioning is performed through step 1041: scanning the color mutation position in the candidate area of the polarization filtering image, and recording it as an edge feature point. Cross-frame trajectory matching: in the previous frame image, the most similar point is searched in the 5x5 pixel range centered on the current edge point: comparing color features, i.e., the similarity of RGB values; comparing shape features, i.e., the gradient pattern of adjacent pixels; selecting the point with the highest similarity as the matching point; vector field construction: calculating the coordinate difference (△x,△y) of the current point and the matching point, and generating an arrow pointing from the position of the previous frame to the position of the current frame: arrow direction = arctan(△y / △x); arrow length = (△x²+△y²)^(1 / 2). For example, the current frame edge point (205,318) is matched to the previous frame (202,315), generating an arrow with a direction of 45° and a length of 4.2 pixels.

[0099] Through step 1042, the starting point coordinate data of each motion vector in the motion vector field is read, i.e., the position information of the current frame edge feature point, and then the pre-stored texture complexity value at the corresponding coordinate position in the texture complexity distribution map is retrieved, establishing a binding relationship table between the motion vector identifier and the texture value. The table is stored in text format as an entry of "motion vector ID, coordinate X, coordinate Y, texture value", for example, "MV102, 355, 422, 1350" represents that the motion vector with identifier MV102 is bound to the texture value 1350 at coordinate (355, 422).

[0100] Through step 1043, based on the texture value in the binding relationship table, the weight factor is calculated through a preset linear conversion coefficient, and the calculation process is: texture value x coefficient, and then a new data structure containing displacement data and weight is created for each motion vector, recorded in the format of "motion vector ID, horizontal displacement, vertical displacement, weight factor", for example, MV102 binds to the value 1350 → weight 1.35 → record as "MV102, 3, 3, 1.35", and finally all records are aggregated in coordinate order to form a comma-separated text collection, such as "MV102, 3, 3, 1.35; MV205, 1, 0, 0.32;...".

[0101] In actual application, in the night highway checkpoint monitoring, through step 1041, the right lower corner edge point of the license plate character after polarization filtering is matched, the current frame coordinate (355, 422) is matched to the best point (352, 419) in the 5x5 pixel search area of the previous frame, and the displacement difference△x=3,△y=3 is calculated to generate a motion vector with a direction of 45 degrees and a length of 4.2 pixels.

[0102] Through step 1042, the starting point coordinate (355, 422) of the vector is read, the same position value 1350 is obtained by querying the texture distribution map, and the binding record "MV102, 355, 422, 1350" is generated.

[0103] The weight 1350x0.001=1.35 is calculated by step 1043, the vector record "MV102, 3, 3, 1.35" is constructed, and the glare area vector is processed, the coordinate (280, 380) texture value 320→weight 0.32→record "MV205, 1, 0, 0.32", and the complete set "MV102, 3, 3, 1.35; MV205, 1, 0, 0.32;..." is output.

[0104] The overall scheme of the above step 104 fuses the polarized enhanced edge feature and the texture anti-interference ability in a weak light environment, constructs a motion vector set with a credibility weight, and provides a reliability quantization basis for subsequent motion analysis, thereby significantly improving the stabilization accuracy of a night motion target.

[0105] 105. Hierarchical clustering the motion vector set according to the texture weight factor, separating a background motion vector group generated by the motion of the unmanned intelligent turntable itself and a foreground abnormal vector group caused by environmental glare interference, and only selecting the geometric center value of the background motion vector group for inter-frame motion compensation to output a stable image sequence.

[0106] Optionally, step 105 can specifically include the following steps:

[0107] 1051. Dividing the motion vector set into a high weight group and a low weight group according to the texture weight factor, the high weight group containing motion vectors with a texture weight factor higher than a preset separation threshold, and the low weight group containing the remaining motion vectors;

[0108] 1052. Calculating the direction consistency density of the motion vectors in the high weight group, and classifying the motion vectors falling into a density peak value region as a background motion vector group, and classifying the low weight group and the motion vectors in the high weight group not falling into the density peak value region as a foreground abnormal vector group;

[0109] 1053. Only selecting the geometric center value of all vectors in the background motion vector group as a translation reference to perform reverse displacement compensation on the next frame of image to output a stable image sequence.

[0110] In the above scheme, the separation threshold is a critical weight value for distinguishing the credibility of the motion vectors. The direction consistency density is the number of concentrated motion vectors in a specific angle range. The density peak value region is a continuous angle interval with the highest direction consistency density. The geometric center value is the average value of the displacement amounts of all background motion vectors. The reverse displacement compensation is a displacement amount opposite to the average displacement amount of the whole image.

[0111] In the embodiments of the present application, first, the weight separation threshold is preset through step 1051, and the motion vector set record is scanned piece by piece, in the format: "ID, horizontal displacement, vertical displacement, weight". When the weight value is greater than or equal to the threshold, the vector is moved into the high weight group, and the rest is moved into the low weight group, generating two independent vector sets. For example, scanning the record "MV102, 3, 3, 1.35; MV205, 1, 0, 0.32"→threshold 1.0→high weight group records MV102, and low weight group records MV205.

[0112] Then, the direction bin statistics are performed through step 1052: 0-360 degrees are divided into 36 direction bins at intervals of 10 degrees, and the high weight group vectors are traversed: the direction angle of each vector is calculated, and the inverse tangent value of the horizontal displacement and the vertical displacement is calculated; the vector count is added to the corresponding direction bin, such as 45.2° into the 40-50° bin; density peak detection: scan the number of vectors in the 36 direction bins, find the highest continuous bin group in the count, for example, 40-50° bin 12, 50-60° bin 11→continuous peak area 40-60°. Group separation: the vectors in the peak area in the high weight group→background group; the non-peak area vectors in the high weight group + all vectors in the low weight group→foreground group. For example, the high weight group contains 20 vectors, of which 15 fall into the 40-60° peak area→the background group collects 15, and the remaining 5 are combined with the 10 vectors in the low weight group to form the foreground group.

[0113] Finally, through step 1053, the background group displacement average: the average value of the horizontal displacement of all background motion vectors and the average value of the vertical displacement are calculated, for example, 10 vectors horizontal displacement and 30 pixels→average value 3 pixels. Reverse translation execution: the next frame image is translated as a whole to the left by the average value of pixels and to the top by the average value of pixels. Sequence output: the translated image is added to the stable video stream, and the original coordinates are superimposed with the displacement compensation value.

[0114] In actual application, in night road monitoring, the motion vector set is input through step 1051: "MV101, 3.1, 3.0, 1.38; MV102, 2.9, 3.2, 1.35; MV201, -1, 4, 0.28; MV301, 0.5, 0.3, 0.41"; the separation threshold is set to 1.0, the high weight group: MV101 (1.38), MV102 (1.35); the low weight group: MV201 (0.28), MV301 (0.41).

[0115] By step 1052, high weight group vector direction calculation: MV101: displacement (3.1, 3.0) -> direction 44.1°; MV102: displacement (2.9, 3.2) -> direction 47.8°; direction binning statistics, 10° / bin: 40-50° bin: 2 vectors; other bins: 0; density peak region: 40-50°, only continuous bin group. Group separation: background group: MV101, MV102, both belong to 40-50° bin; foreground group: MV201 direction 104°, MV301 direction 31°.

[0116] By step 1053, background group displacement average: average = (3.1 + 2.9) / 2 = 3.0 pixels; vertical average = (3.0 + 3.2) / 2 = 3.1 pixels; reverse compensation: next frame image left shift 3.0 pixels, up shift 3.1 pixels; output verification: after compensation, license plate center coordinate fluctuation range: ±0.2 pixels, original ±3.5 pixels; no ghosting on license plate character edge. The overall scheme of the above step 105, through texture weight layering and direction density peak analysis, accurately separates the real motion component of the turntable in a weak light interference environment, performs image compensation based on physically consistent motion average, completely eliminates picture jitter caused by glare artifacts in night monitoring, and outputs stable and usable video sequences.

[0117] The following is a complete embodiment for steps 101-105:

[0118] In night road monitoring, by step 101, an unmanned intelligent turntable tracks a suspected vehicle at a speed of 30° / second, synchronously acquiring continuous image frames interfered by strong glare from car lights, with a frame rate of 50fps. Identification process: divide each image into 16x16 pixel units, calculate the maximum and minimum brightness difference in the unit; mark units with a difference of <25 as low texture units; merge adjacent low texture units to generate a texture blur area covering the front bumper and road reflection strip.

[0119] By step 102, for the bumper texture blur area, perform adjacent frame structure change calculation: take the current frame and the frame 0.02 seconds ago, and accumulate the absolute value of pixel brightness difference in each unit; generate a texture complexity distribution map. High value area: 1350 at license plate, low value area: 280 at car light direct reflection; set threshold 1000, merge continuous high value units around the license plate to form a "U" shaped candidate area with an area of 120x40 pixels, which clearly retains the license plate character indentation and screw hole texture.

[0120] Through step 103, the candidate region is analyzed for continuous 5 frame pixel brightness fluctuation: license plate character point brightness sequence [50, 120, 55, 125, 60] → dominant polarization direction 0°; rotate the light transmission axis of the filter to 90°, the current angle is 170° → rotate counterclockwise by 80°; collect the image after filtering: the brightness of the car light glare decreases from 200 to 60, the grayscale transition of the license plate edge changes from [30, 160] to [65, 68, 71], and the exposure time is compensated to 1 / 67 second, originally 1 / 100 second.

[0121] Through step 104, in the license plate candidate area of the polarization filtering image: track the right lower corner point of the license plate character: the current frame (355, 422) matches the previous frame (352, 419) → displacement (3, 3) → vector 45° direction / 4.2 pixel motion vector (355, 422) binds texture value 1350 → weight = 1350 * 0.001 = 1.35 Glare point vector (280, 380) binds texture value 320 → weight 0.32 Output set: "MV102, 3, 3, 1.35; MV205, 1, 0, 0.32".

[0122] Through step 105, weight stratification: threshold 1.0 → high weight group MV102 (1.35), low weight group MV205 (0.32) Direction density statistics: high weight group vector direction 45° & 46.8° → 40-50° box density highest Background group: MV102 falls into the peak area; geometric center value: horizontal displacement 3 pixels, vertical displacement 3 pixels Reverse compensation: next frame left shift 3 pixels / up shift 3 pixels Output result: license plate center coordinate fluctuation decreases from ±3.5 pixels to ±0.2 pixels.

[0123] In the night road monitoring scene with weak light intensity glare, the scheme realizes high-precision image stabilization of the moving target through the threefold synergistic mechanism of texture anti-interference capability quantification screening, dynamic polarization filtering to enhance edge features, and texture weight constraint motion vector clustering, accurately separates the real motion of the turntable from the environmental optical artifacts, and realizes high-precision image stabilization of the moving target. The output image sequence completely eliminates the edge tearing and high-frequency jitter caused by car light glare, and significantly improves the reliability of suspicious vehicle tracking and identification.

[0124] Figure 3 An embodiment of the present application provides a structural schematic diagram of a real-time image stabilization system for intelligent image processing, as shown in Figure 3 The system comprises:

[0125] The acquisition module 31 is configured to, when the unmanned intelligent turntable performs pitching or rotating motion, synchronously acquire continuous image frames disturbed by environmental strong glare, and identify a texture blurred area caused by light reflection in the continuous image frames.

[0126] The computing module 32 is configured to calculate a texture complexity distribution map in combination with a local structure change amount of an adjacent image frame for the texture blur area, and screen a candidate area with a texture feature retention rate higher than a preset standard in the texture blur area based on the texture complexity distribution map.

[0127] The processing module 33 is configured to apply a polarized light filtering process to the candidate area, generate a polarized light filtering image by adjusting a filtering angle to match an incident light polarization direction of a current frame, and the polarized light filtering image has both a glare noise suppression feature and an edge gradient continuity feature.

[0128] The extraction module 34 is configured to extract a motion vector field of the candidate area between adjacent frames based on the gradient continuity feature of the polarized light filtering image, and perform spatial superposition of the motion vector field and the texture complexity distribution map to form a motion vector set carrying a texture weight factor.

[0129] The separation module 35 is configured to perform hierarchical clustering on the motion vector set according to the texture weight factor, separate a background motion vector group generated by self-motion of the unmanned intelligent turntable and a foreground abnormal vector group caused by environmental glare interference, and only select a geometric center value of the background motion vector group for inter-frame motion compensation to output a stable image sequence.

[0130] Figure 3 The real-time image stabilization system for intelligent image processing can perform Figure 1 The real-time image stabilization method for intelligent image processing in the embodiment has the implementation principle and technical effects which will not be repeated. The specific operation manner of each module and unit in the real-time image stabilization system for intelligent image processing in the above embodiment has been described in detail in the embodiment related to the method, and will not be described in detail here.

[0131] In one possible design, Figure 3 The real-time image stabilization system for intelligent image processing in the embodiment can be implemented as a computing device, such as a server or a terminal. Figure 4 As shown in the figure, the computing device can include a storage component 41 and a processing component 42.

[0132] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 42.

[0133] The processing component 42 is configured to perform the above Figure 1 The real-time image stabilization method for intelligent image processing in the embodiment.

[0134] The processing component 42 can include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component can also be one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, micro-controllers, microprocessors or other electronic components, configured to perform the methods described above.

[0135] The storage component 41 is configured to store various types of data to support the operation of the terminal. The storage component can be implemented by any type of volatile or non-volatile storage devices or their combinations, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0136] Of course, the computing device can also include other components, such as an input / output interface, a display component, a communication component, etc.

[0137] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0138] The communication component is configured to facilitate wired or wireless communication between the computing device and other devices, etc.

[0139] The computing device can be a physical device or an elastic computing host provided by a cloud computing platform, and the computing device can be a cloud server, and the processing component, the storage component, etc. can be a basic server resource rented or purchased from the cloud computing platform.

[0140] The embodiment of the present application also provides a computer storage medium, which stores a computer program, and the computer program can implement the above Figure 1 A real-time image stabilization method for intelligent image processing is provided in the embodiment.

[0141] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, device and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0142] The apparatus embodiments described above are merely illustrative, wherein the units described as separated components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., can be located in one place, or can be distributed to multiple network units. Part or all of the modules can be selected to achieve the purpose of the embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0143] Through the description of the above embodiments, those skilled in the art can clearly understand that the embodiments can be realized by means of software and the necessary general hardware platform, and of course can also be realized by hardware. Based on such understanding, the above technical solutions can be embodied in the form of software products, and the computer software products can be stored in a computer readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and include a plurality of instructions to make a computer device (which can be a personal computer, a server, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0144] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A real-time image stabilization method for intelligent image processing, characterized in that, The application relates to a method for generating a stable image sequence by eliminating the interference of strong ambient light on a continuous image frame. When the unmanned intelligent rotary table performs pitching or rotating motion, a continuous image frame interfered by strong ambient light is synchronously acquired, and a texture blurred area caused by light reflection in the continuous image frame is identified; For the texture blurred area, a texture complexity distribution graph is calculated in combination with the local structure change amount of adjacent image frames, and a candidate area with a texture feature retention rate higher than a preset standard is screened based on the texture complexity distribution graph; Polarization filtering processing is applied to the candidate area, a polarization filtering image is generated by adjusting a filtering angle to match the incident light polarization direction of the current frame, and the polarization filtering image has the gradient continuity features of suppressing glare noise and enhancing edge gradient; Based on the gradient continuity features of the polarization filtering image, a motion vector field of the candidate area between adjacent frames is extracted, and the motion vector field is spatially superimposed with the texture complexity distribution graph to form a motion vector set carrying a texture weight factor; According to the texture weight factor, the motion vector set is hierarchically clustered to separate a background motion vector group generated by the motion of the unmanned intelligent rotary table and a foreground abnormal vector group caused by the interference of ambient light, only the geometric center value of the background motion vector group is selected for inter-frame motion compensation, and a stable image sequence is output.

2. The method of claim 1, wherein, Based on the gradient continuity features of the polarization filtering image, a motion vector field of the candidate area between adjacent frames is extracted, and the motion vector field is spatially superimposed with the texture complexity distribution graph to form a motion vector set carrying a texture weight factor, which comprises: In the candidate area of the polarization filtering image, the displacement trajectory of the corresponding edge points between adjacent frames is tracked along the object edge direction, and a motion vector field is formed by the direction and length of the displacement trajectory; The spatial position of each motion vector in the motion vector field is bound with the value of the texture complexity distribution graph at the same spatial position; According to the bound value, each motion vector is given a texture weight factor, and the texture weight factor is in a proportional relationship with the value of the texture complexity distribution graph, thereby forming a motion vector set carrying a texture weight factor.

3. The method of claim 1, wherein, The polarization filtering processing applied to the candidate area generates a polarization filtering image by adjusting a filtering angle to match the incident light polarization direction of the current frame, and the polarization filtering image has the gradient continuity features of suppressing glare noise and enhancing edge gradient, which comprises: According to the light reflection characteristics of the candidate area, the dominant polarization direction of the incident light in the candidate area is analyzed, and the dominant polarization direction is determined by the periodic change mode of the pixel brightness in the continuous frames; The light transmission axis angle of the physical filter is dynamically rotated to form an orthogonal relationship between the light transmission axis angle and the dominant polarization direction; According to the adjusted light transmission axis angle, the light rays of the candidate area passing through the physical filter are collected to generate a polarization filtering image, and the brightness amplitude of the glare noise in the polarization filtering image is reduced and the light-dark transition continuity of the object edge is improved.

4. The method of claim 1, wherein, The motion vector set is hierarchically clustered according to the texture weight factor, background motion vector groups generated by self-motion of the unmanned intelligent turntable and foreground abnormal vector groups caused by environmental glare interference are separated, only the geometric center value of the background motion vector groups is selected for inter-frame motion compensation, and a stable image sequence is output, including: The motion vector set is divided into a high weight group and a low weight group according to the texture weight factor, the high weight group contains motion vectors with a texture weight factor higher than a preset separation threshold, and the low weight group contains the remaining motion vectors; The direction consistency density of the motion vectors in the high weight group is calculated, the motion vectors falling into the density peak value region are classified as background motion vector groups, and the low weight group and the motion vectors not falling into the density peak value region in the high weight group are classified as foreground abnormal vector groups; Only the geometric center value of all vectors in the background motion vector groups is selected as the translation reference, reverse displacement compensation is performed on the next frame of image, and a stable image sequence is output.

5. The method of claim 1, wherein, For the texture blur area, a texture complexity distribution map is calculated in combination with the local structure change amount of the adjacent image frames, and a candidate area with a texture feature retention rate higher than a preset standard in the texture blur area is selected based on the texture complexity distribution map, including: For each image unit in the texture blur area, the structure change amount of the corresponding image unit between the current frame and the adjacent frame is calculated, and the structure change amount is represented by the cumulative value of the difference amplitude of the pixel values in the unit; The structure change amount is mapped to each image unit position of the texture blur area to generate a texture complexity distribution map, and the value of the texture complexity distribution map reflects the ability strength of the image unit to resist glare interference; In the texture complexity distribution map, image units with a value higher than a preset threshold are marked, and continuously distributed marked units in the image units are combined into a candidate area, and the candidate area covers a continuous spatial range with a texture feature retention rate higher than a preset standard.

6. The method of claim 3, wherein, The light transmission axis angle of the dynamic rotating physical filter is adjusted to form an orthogonal relationship between the light transmission axis angle and the dominant polarization direction, including: The angle value of the dominant polarization direction is measured, and the orthogonal angle gap between the angle value and the light transmission axis angle is calculated; The motor driving pulse number is generated according to the orthogonal angle gap, and the rotating motor is driven to rotate the physical filter by the gap compensation angle; During rotation, when the feedback value of the light transmission axis angle sensor reaches the orthogonal angle of the angle value, the rotating motor is stopped and the light transmission axis is locked.

7. The method of claim 2, wherein, According to the bound values, a texture weight factor is assigned to each motion vector, the texture weight factor is in a positive proportional relationship with the value of the texture complexity distribution map, and a motion vector set carrying the texture weight factor is formed, including: The bound values and the texture complexity distribution map values corresponding to the motion vector spatial positions are taken as the texture credibility values of the current motion vectors; The texture credibility value is input into a linear converter to generate a texture weight factor through a fixed-multiplying amplification operation, and a value of the texture weight factor is equal to the texture credibility value multiplied by a preset proportion coefficient; The texture weight factor is written into an attribute field of a corresponding motion vector to construct a motion vector set carrying the texture weight factor.

8. A real-time image stabilization system for intelligent image processing, characterized in that, The method comprises the following steps: The acquisition module is configured to acquire, synchronously, continuous image frames disturbed by environmental strong glare light when an unmanned intelligent turntable performs pitching or rotating motion, and identify a texture blur area caused by light reflection in the continuous image frames; The calculation module is configured to calculate a texture complexity distribution graph in combination with a local structure change amount of adjacent image frames for the texture blur area, and filter a candidate area with a texture feature retention rate higher than a preset standard based on the texture complexity distribution graph; The processing module is configured to apply a polarized light filtering process to the candidate area, generate a polarized light filtering image by adjusting a filtering angle to match an incident light polarization direction of a current frame, and the polarized light filtering image has both glare noise suppression and gradient continuity features of enhanced edge gradient; The extraction module is configured to extract a motion vector field of the candidate area between adjacent frames based on the gradient continuity features of the polarized light filtering image, and perform spatial superposition of the motion vector field and the texture complexity distribution graph to form a motion vector set carrying a texture weight factor; The separation module is configured to perform hierarchical clustering on the motion vector set according to the texture weight factor, separate a background motion vector group generated by the motion of the unmanned intelligent turntable itself and a foreground abnormal vector group caused by environmental glare light disturbance, select only a geometric center value of the background motion vector group for inter-frame motion compensation, and output a stable image sequence.

9. A computing device, comprising: The method comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement the real-time image stabilization method for intelligent image processing according to any one of claims 1 to 7.

10. A computer storage medium, characterized in that, The computer program is stored in a computer and is executed by the computer to implement the real-time image stabilization method for intelligent image processing according to any one of claims 1 to 7.

Images