Image anti-shake method and system for unmanned aerial vehicle inspection

Abstract

The application relates to the technical field of image communication transmission, in particular to an image anti-shake method and system for unmanned aerial vehicle inspection. The method comprises the following steps: acquiring a video stream collected by an unmanned aerial vehicle, and dividing each frame image of the video stream into multiple grids; for any grid in any frame image, calculating a gradient oscillation index and a turbulent motion entropy of the grid, fusing the gradient oscillation index and the turbulent motion entropy to obtain a gas interference index; determining an attenuation factor corresponding to the grid based on the gas interference index; weighting a data fidelity term of an objective function in an L1 norm optimization algorithm by using the attenuation factor, processing the grid by using the L1 norm optimization algorithm based on the weighted data fidelity term to obtain grid vertex deformation parameters; and performing image stabilization processing on the video stream according to the grid vertex deformation parameters. The application can reduce the local geometric distortion problem caused by a traditional anti-shake algorithm, and improve the fidelity of an image after mechanical jitter is eliminated.

Description

Technical Field

[0001] This application relates to the field of image communication transmission technology, and in particular to an image stabilization method and system for unmanned aerial vehicle (UAV) inspection. Background Technology

[0002] In the petrochemical industry, using drones equipped with optical gas imaging (OGI) devices to inspect pipelines and equipment has become a mainstream method for monitoring gas leaks, replacing high-risk manual operations. However, when drones operate in complex chemical industrial park environments, the video images they acquire are often subject to severe shaking and blurring due to interference from omnidirectional airflow at high altitudes and high-frequency vibrations of the drone's rotor. This shaking is particularly noticeable when capturing tiny leaks with a telephoto lens, which can disrupt the visual characteristics of gas plumes, making it difficult to locate leaks and even causing missed detections or misjudgments by the monitoring system.

[0003] Among related video stabilization technologies, the most commonly used is the grid-based motion stabilization with L1 optimization algorithm. This algorithm divides the image into several grids, calculates the optical flow trajectory of feature points, and uses the L1 norm to optimize the objective function to smooth the camera path. It has advantages such as preserving the image's geometric structure and strong ability to adapt to parallax changes.

[0004] However, in the specific scenario of inspecting petrochemical gas leaks, the images not only contain a static, rigid background but also a large number of non-rigid moving objects, such as the high-speed flowing gas from the leak and heat waves generated by high-temperature equipment. Traditional L1 norm optimization algorithms, when processing video frames, struggle to distinguish between the mechanical jitter of the drone and the local fluid motion of the gas and heat waves, easily misclassifying gas flow as background motion and including it in the optimization solution. This deficiency leads to inaccurate calculated mesh deformation parameters, resulting in local geometric distortion in the final stabilized video, causing image distortion. Summary of the Invention

[0005] To address the issue of local geometric distortion that occurs after image stabilization and to improve image fidelity, this application provides an image stabilization method and system for UAV inspection.

[0006] Firstly, this application provides an image stabilization method for drone inspection, employing the following technical solution:

[0007] Image stabilization method for UAV inspection includes: acquiring a video stream captured by the UAV and dividing each frame of the video stream into multiple grids;

[0008] For any grid in any frame of image, calculate its gradient oscillation index and turbulent motion entropy. The gradient oscillation index is used to characterize the degree of gradient change of the pixels in the grid within a preset time window, and the turbulent motion entropy is used to characterize the consistency of the motion vector of each pixel in the grid.

[0009] The gas disturbance index is obtained by fusing the gradient oscillation index and the turbulent motion entropy. The gas disturbance index is positively correlated with the gradient oscillation index and the turbulent motion entropy.

[0010] The attenuation factor corresponding to the grid is determined based on the gas disturbance index, wherein the attenuation factor is negatively correlated with the gas disturbance index;

[0011] The data fidelity term of the objective function in the L1 norm optimization algorithm is weighted by an attenuation factor. Based on the weighted data fidelity term, the L1 norm optimization algorithm is used to process the mesh to obtain the mesh vertex deformation parameters. The video stream is then stabilized based on the mesh vertex deformation parameters.

[0012] In video streams, gas regions typically correspond to large gradient changes and chaotic pixel motion. Therefore, this application introduces two features: gradient oscillation index and turbulent motion entropy. These two indicators comprehensively characterize the temporal instability of gradients and the consistency of motion vectors in the image grid region, thereby accurately identifying gas or heat wave interference regions in the image. Based on this, by constructing a decay factor negatively correlated with the gas interference index, the data fidelity term of the L1 norm optimization objective function is adaptively weighted, reducing the weight of gas regions in the image stabilization solution, making the optimization process rely solely on the rigid background calculation path. This not only effectively removes image jitter but also avoids plume morphology distortion caused by forcibly straightening gas motion, improving image fidelity and the accuracy of subsequent leak point location.

[0013] Optionally, calculating the gradient oscillation index includes: constructing a temporal analysis window for any frame of the image; constructing a local volatility for any pixel within the grid based on the difference in gradient magnitudes between adjacent consecutive frames in the temporal analysis window; and using the sum of the local volatility corresponding to all pixels within the grid as the gradient oscillation index.

[0014] The Schlieren Effect, generated by leaked high-pressure gas or high-temperature steam, causes high-frequency, non-rigid fluctuations in the pixel gradient of the imaging plane. Therefore, by constructing a time analysis window and calculating the difference in gradient magnitude between adjacent consecutive frames, the instability of pixel grayscale in the time domain can be reflected.

[0015] Optionally, the steps for calculating the turbulent motion entropy include: obtaining the motion vector of each pixel in the grid using an optical flow estimation algorithm; calculating the directional information entropy based on the directional distribution of the motion vectors; obtaining the coefficient of variation of the magnitude distribution of the motion vectors; and using the product of the directional information entropy and the coefficient of variation as the turbulent motion entropy.

[0016] By leveraging the highly disordered nature of turbulent fluid motion inherent in gas leakage, and the highly consistent differences in background motion caused by drone jitter, this difference is quantified by calculating the directional information entropy and the coefficient of variation of the motion vector. The directional information entropy identifies regions with chaotic motion directions, while the coefficient of variation identifies regions with large velocity fluctuations. Their product, the turbulent motion entropy, can be used to further confirm and eliminate regions affected by motion blur caused by rapid camera rotation at the kinematic feature level, thereby improving the accuracy of identifying non-rigid fluid interference zones.

[0017] Optionally, the step of calculating the directional information entropy based on the directional distribution of the motion vector includes: dividing the range [0°, 360°] into a preset number of directional intervals, counting the frequency of the motion vector direction of the pixels in the grid falling into each interval, and calculating its information entropy.

[0018] By dividing the 360-degree range into a preset number of intervals and calculating the information entropy by statistically analyzing the frequency, it is possible to effectively capture the randomness of the movement of gas molecules or heat waves in various directions. The entropy value is the largest when the movement direction of the pixel is evenly distributed in multiple directions, thus keenly capturing turbulence characteristics.

[0019] Optionally, the range [0°, 360°] can be divided into 12 directional intervals.

[0020] Optionally, the step of obtaining the coefficient of variation of the magnitude distribution of motion vectors includes: using the ratio of the standard deviation to the mean of the magnitude of the motion vector of each pixel as the coefficient of variation.

[0021] By calculating the ratio of the standard deviation to the mean of the motion vector magnitude as the coefficient of variation, this application normalizes the measurement of velocity fluctuation and eliminates the influence of the absolute magnitude of the motion vector on the judgment of fluctuation.

[0022] Optionally, the step of fusing the gradient oscillation index and the turbulent motion entropy to obtain the gas disturbance index includes: adding the turbulent motion entropy to a preset value to obtain an adjustment factor, and adding the adjustment factor to the gradient oscillation index to obtain the gas disturbance index.

[0023] By combining the gradient oscillation exponent and the turbulent motion entropy through addition and introducing a preset value for adjustment, it is possible to achieve complementarity of multidimensional features and improve the flexibility of the system.

[0024] Optionally, the step of determining the attenuation factor corresponding to the grid based on the gas disturbance index includes: normalizing the gas disturbance index; the formula for calculating the attenuation factor is:

[0025] In the formula, Indicates the first The attenuation factor of each grid; Indicates the first Normalized gas disturbance index for each grid; Represented by natural constant An exponential function with base 0.

[0026] When the gas interference index of the grid area is high, it indicates that there is gas interference in the area, and the attenuation factor quickly approaches 0, achieving strong shielding; when it is determined to be a rigid background, the attenuation factor approaches 1, achieving fidelity preservation for the area.

[0027] Optionally, in the step of acquiring the video stream captured by the drone, the video stream is a pre-processed video stream, and the pre-processing operation includes at least one of grayscale processing, adaptive median filtering processing, and histogram equalization processing.

[0028] Secondly, this application provides an image stabilization system for drone inspection, which adopts the following technical solution:

[0029] An image stabilization system for drone inspection includes a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the image stabilization method for drone inspection described above is implemented.

[0030] The image stabilization method described above for UAV inspection is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. Thus, a system can be built based on the memory and processor for easy use.

[0031] This application has the following technical advantages:

[0032] A gradient oscillation index characterizing temporal instability and a turbulent motion entropy characterizing motion disorder are constructed to distinguish between gas regions and rigid background regions. A decay factor is generated by fusing these two indices, and the data fidelity term of the L1 optimization objective function is adaptively weighted. Non-rigid fluid interference is precisely shielded, and a stable path is calculated using only a rigid background, thus eliminating jitter while preserving the true shape of the gas plume. Attached Figure Description

[0033] Figure 1 This is a flowchart of an image stabilization method for drone inspection according to an embodiment of this application.

[0034] Figure 2 These are video frames captured by a drone.

[0035] Figure 3 This is a diagram showing the effect of processing video frames captured by a drone using a traditional L1 norm optimized anti-shake algorithm.

[0036] Figure 4 This is a diagram showing the effect of the L1 norm optimized anti-shake algorithm in this application on video frames captured by the drone. Detailed Implementation

[0037] This application discloses an image stabilization method for UAV inspection. Addressing the problem in petrochemical scenarios where mechanical shaking and non-rigid motion from gas leaks are difficult to distinguish in UAV inspection videos, leading to local geometric distortion in traditional stabilization algorithms, this application proposes an improved L1 norm-based stabilization scheme weighted by multi-dimensional feature confidence. By constructing a gradient oscillation index, turbulent motion entropy, and attenuation factor, it accurately identifies and masks gas / heat wave interference areas, calculating the image stabilization path using only rigid background features. This removes shaking while maintaining the true shape of the gas plume in the image, reducing the occurrence of local distortion.

[0038] Reference Figure 1 The image stabilization method for drone inspection includes steps S1-S5.

[0039] S1: Acquire the video stream captured by the drone and divide each frame of the video stream into multiple grids.

[0040] In the inspection scenario of a petrochemical industrial park, a drone equipped with an optical gas imaging (OGI) sensor and an inertial measurement unit (IMU) flies along a preset route. The OGI sensor operates in a specific infrared band (e.g., This band exhibits absorption characteristics for hydrocarbon gases. Simultaneously, to ensure a one-to-one correspondence between the thermal imaging characteristics of the gas leak point and the visible light environment characteristics, the acquisition system uses a hardware synchronization trigger signal to ensure frame-level synchronization between the infrared thermal imaging video stream and the aircraft attitude angle data.

[0041] The acquired raw infrared video stream is preprocessed, including: converting each frame of the raw image to grayscale to reduce computational dimensionality; employing adaptive median filtering to denoise thermal noise generated by high-temperature equipment, removing salt-and-pepper noise while preserving equipment edge details; and then performing histogram equalization to stretch the grayscale distribution, enhancing the contrast between the gas plume and the background sky or pipes.

[0042] Finally, each preprocessed frame image is divided into... A regular grid is used as the basic unit for subsequent motion analysis. For example, the number of grid cells can be set according to the image resolution, such as for... Images of resolution can be divided into: Each grid.

[0043] S2: For any grid in any frame of image, calculate its gradient oscillation index and turbulent motion entropy. The gradient oscillation index is used to characterize the degree of gradient change of the pixels in the grid within a preset time window, and the turbulent motion entropy is used to characterize the consistency of the motion vector of each pixel in the grid.

[0044] Rapid turbulent gas flow and local refractive index changes cause high-frequency, non-rigid, and violent fluctuations in the pixel gradient of the corresponding region on the imaging plane. Therefore, in chemical inspection scenarios, leaking high-pressure gas or accompanying high-temperature steam can produce a Schlieren effect, resulting in high-frequency, non-rigid fluctuations (i.e., flickering) in the pixel gradient of the corresponding region on the imaging plane. In contrast, rigid backgrounds, such as pipes and towers, although displaced with the drone's movement, maintain a relatively stable texture gradient structure. Based on this difference, a gradient oscillation index is constructed to quantify the instability of temporal gradients.

[0045] Specifically, the steps for calculating the gradient oscillation exponent include: for any frame of image, extracting the gradient magnitude sequence of each pixel within a time analysis window. In this embodiment, a time analysis window is constructed based on the frame of image, and the length of the time analysis window is [missing information]. In this embodiment The length is 10, which means that the time analysis window includes the 10 consecutive frames before the current frame. For any grid, the gradient magnitude of the pixels in the grid is extracted in the time analysis window to obtain the gradient magnitude sequence.

[0046] For any grid, the formula for calculating its gradient oscillation exponent can be expressed as:

[0047] In the formula, Indicates the first Gradient oscillation index of each grid; This indicates the total number of pixels within the grid. The length of the time analysis window; Indicates the first In the grid, the th The pixel in the time analysis window Time frame and the The maximum gradient magnitude between time frames; Indicates the first In the grid, the th The pixel in the time analysis window Time frame and the The minimum value of the gradient magnitude between time frames; This is a preset parameter, and in this embodiment, it is set to 1, which is mainly used to prevent the denominator from being zero.

[0048] This metric calculates the cumulative relative rate of change of the gradient of each pixel within the grid. If there is a gas leak within the grid area, thermal and optical disturbances cause the pixel edge gradient values ​​to fluctuate wildly over time, resulting in an increase in the numerator. As the value of increases, the gradient oscillation exponent increases after being amplified by the squared term. If the mesh is a rigid background, even with translation, its gradient magnitude remains essentially unchanged, and the numerator term is close to zero, resulting in a smaller gradient oscillation exponent value. Thus, rigid objects and non-rigid fluids can be preliminarily distinguished based on the temporal instability of the gradient.

[0049] Relying solely on gradient oscillations may be affected by motion blur caused by rapid camera rotation; therefore, secondary verification is necessary by incorporating the kinematic characteristics of the optical flow vector field. Gas leakage is a type of fluid turbulence, and its motion vector field exhibits high disorder, while the background motion vector caused by the rapid rotation of the UAV shows high consistency.

[0050] The steps for calculating the turbulent motion entropy include: for any grid in any frame of image, using an optical flow field estimation algorithm to obtain the motion vector of each pixel in that grid. Taking the first frame image as an example, the first frame... Frame image and the first The first frame of the image The gray values ​​of pixels within a grid region are used as input to the Farneback dense optical flow estimation algorithm. The Farneback algorithm is based on polynomial expansion. It approximates the image neighborhood with a quadratic polynomial and estimates the displacement using the transformation relationship of the polynomial coefficients. It outputs the motion vector of all pixels within the grid region. The Farneback dense optical flow estimation algorithm is a conventional technique in this field, so it will not be described in detail here.

[0051] After obtaining the motion vectors of each pixel in the grid region, the turbulent motion entropy can be calculated based on the motion vectors.

[0052] In the process of calculating the entropy of turbulent motion, firstly... The direction is evenly divided into a predetermined number of intervals; in this embodiment, it is divided into 12 intervals to balance computational efficiency and accuracy. The first interval is then statistically analyzed. The information entropy is calculated by calculating the frequency of the motion vector directions of all pixels within each grid interval. Then, based on the... The coefficient of variation is calculated by dividing the standard deviation of the motion vector magnitude of all pixels within a grid by the mean. It reflects the degree of fluctuation in the magnitude of speed.

[0053] For any given grid, the product of the information entropy and the coefficient of variation corresponding to that grid is taken as the turbulent motion entropy. The formula for calculating the turbulent motion entropy can be expressed as:

[0054] In the formula, Indicates the first The turbulent motion entropy of each grid; Indicates the first The directional information entropy of each grid; Indicates the first The coefficient of variation for each grid cell.

[0055] Gas leaks involve turbulent fluid motion, resulting in inconsistent motion directions at different points within the leak area. This leads to a uniform directional distribution and high information entropy; simultaneously, the motion vector fluctuates significantly. Therefore, through... This can further improve the accuracy of identifying gas leak areas.

[0056] S3: The gas disturbance index is obtained by fusing the gradient oscillation index and the turbulent motion entropy. The gas disturbance index is positively correlated with the gradient oscillation index and the turbulent motion entropy.

[0057] A gas disturbance index is constructed based on turbulent entropy and gradient oscillation index. In this embodiment, the sum of a preset value and turbulent entropy is used as an adjustment factor, and the sum of the adjustment factor and gradient oscillation index is used as the gas disturbance index. In this embodiment, the preset value is 1.

[0058] By constructing a gas interference index using synergistic turbulent motion entropy and gradient oscillation index, regions with consistent motion vectors despite gradient changes are accurately eliminated. These regions are rigid background areas and not gas leaks, thus further improving the identification accuracy of non-rigid gas interference regions.

[0059] S4: Determine the attenuation factor corresponding to the grid based on the gas interference index, where the attenuation factor is negatively correlated with the gas interference index.

[0060] In L1 norm-optimized image stabilization algorithms, the goal is to find a smooth camera path, with the core assumption that mesh deformation should preserve geometric structure. However, mesh deformation in gas regions is physically real fluid deformation, not camera shake. Forcing optimization of these meshes would distort the entire image to straighten gas movement, leading to geometric distortion. Therefore, regions with more pronounced gas characteristics carry less camera pose information, resulting in lower reliability. Thus, a normalized attenuation factor is constructed in this step to shield the influence of gas regions in subsequent optimization equations.

[0061] First, the gas disturbance index of all grids in the entire map is collected and processed using the max-min normalization method. Then, an attenuation factor is constructed based on the normalized gas disturbance index. For any grid, the formula for calculating the corresponding attenuation factor can be expressed as:

[0062] In the formula, Indicates the first The attenuation factor for each grid cell ranges from 1 to 1. ; Indicates the first Normalized gas disturbance index for each grid; Represented by natural constant An exponential function with base 0.

[0063] As can be seen from this formula, when the grid is a gas interference zone, Larger, approaching 1 As the exponential function increases, the attenuation factor rapidly approaches 0. When the mesh has a rigid background, As the denominator approaches 0, the denominator approaches 1, leading to... Close to 1.

[0064] This nonlinear mapping adaptively assigns weights to the data fidelity terms of each grid, achieving strong suppression of high-intensity gas disturbances.

[0065] S5: The data fidelity term of the objective function in the L1 norm optimization algorithm is weighted by an attenuation factor. Based on the weighted data fidelity term, the L1 norm optimization algorithm is used to process the mesh to obtain the mesh vertex deformation parameters. The video stream is then stabilized according to the mesh vertex deformation parameters.

[0066] A grid-based L1 norm-optimized anti-shake algorithm is employed, whose original objective function typically includes a data fidelity term and a smoothness term. In this embodiment, the attenuation factor obtained in step S4 is used to perform spatial adaptive weighting on the data fidelity term.

[0067] when When the value approaches 0, it indicates that the region is a gas region, and the contribution of this mesh to the objective function becomes smaller. This causes the optimizer to automatically ignore the disordered motion information in these regions when solving for the global camera path, and only rely on the more weighted pipe and ground-like rigid background meshes to calculate the camera path.

[0068] After completing the weighted L1 optimization, a set of mesh vertex deformation parameters is obtained that only reflects the actual camera shake and eliminates gas turbulence interference. These parameters are then used to perform a reverse warping transformation on the original video frames to generate a stable video sequence. (Refer to...) Figure 2 , Figure 3 and Figure 4 , Figure 2 The middle part is a video frame captured by a drone. Figure 3 It is a traditional algorithm Figure 2 After image stabilization, the traditional L1 norm optimized image stabilization algorithm loses local structural details when processing video frames captured by the drone, resulting in local blurring and distortion of the image. Figure 4 This is a rendering of the L1 norm optimized anti-shake algorithm in this application. In the rendering, the background pipe remains stable, eliminating mechanical shaking, while the gas plume retains its natural flow state. This reduces gas shape distortion or background stretching caused by algorithm misjudgment and ensures the accuracy of subsequent leak point location.

[0069] This application also discloses an image stabilization system for drone inspection, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the image stabilization method for drone inspection according to this application.

[0070] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0071] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. An image stabilization method for unmanned aerial vehicle (UAV) inspection, characterized in that, Acquire the video stream captured by the drone and divide each frame of the video stream into multiple grids; For any grid in any frame of image, calculate its gradient oscillation index and turbulent motion entropy. The gradient oscillation index is used to characterize the degree of gradient change of the pixels in the grid within a preset time window, and the turbulent motion entropy is used to characterize the consistency of the motion vector of each pixel in the grid. The gas disturbance index is obtained by fusing the gradient oscillation index and the turbulent motion entropy. The gas disturbance index is positively correlated with the gradient oscillation index and the turbulent motion entropy. The attenuation factor corresponding to the grid is determined based on the gas disturbance index, wherein the attenuation factor is negatively correlated with the gas disturbance index; The data fidelity term of the objective function in the L1 norm optimization algorithm is weighted by an attenuation factor. Based on the weighted data fidelity term, the L1 norm optimization algorithm is used to process the mesh to obtain the mesh vertex deformation parameters. The video stream is then stabilized based on the mesh vertex deformation parameters.

2. The image stabilization method for UAV inspection according to claim 1, characterized in that, The calculation of the gradient oscillation index includes: constructing a time analysis window for any frame of the image; constructing a local volatility for any pixel within the grid based on the difference in gradient magnitude between adjacent consecutive frames in the time analysis window; and using the sum of the local volatility corresponding to all pixels within the grid as the gradient oscillation index.

3. The image stabilization method for UAV inspection according to claim 1, characterized in that, The steps for calculating the turbulent motion entropy include: obtaining the motion vector of each pixel in the grid using an optical flow estimation algorithm; calculating the directional information entropy based on the directional distribution of the motion vectors; obtaining the coefficient of variation of the magnitude distribution of the motion vectors; and using the product of the directional information entropy and the coefficient of variation as the turbulent motion entropy.

4. The image stabilization method for UAV inspection according to claim 3, characterized in that, The steps for calculating the directional information entropy based on the directional distribution of the motion vectors include: dividing the range [0°, 360°] into a preset number of directional intervals, counting the frequency of the motion vector direction of the pixels in the grid falling into each interval, and calculating its information entropy.

5. The image stabilization method for UAV inspection according to claim 4, characterized in that, The range [0°, 360°] is divided into 12 directional intervals.

6. The image stabilization method for UAV inspection according to claim 3, characterized in that, The steps to obtain the coefficient of variation of the magnitude distribution of motion vectors include: using the ratio of the standard deviation to the mean of the magnitude of the motion vector of each pixel as the coefficient of variation.

7. The image stabilization method for UAV inspection according to claim 1, characterized in that, The steps for fusing the gradient oscillation index and the turbulent motion entropy to obtain the gas disturbance index include: adding the turbulent motion entropy to a preset value to obtain an adjustment factor, and adding the adjustment factor to the gradient oscillation index to obtain the gas disturbance index.

8. The image stabilization method for UAV inspection according to claim 1, characterized in that, The steps for determining the attenuation factor corresponding to the grid based on the gas disturbance index include: normalizing the gas disturbance index; and calculating the attenuation factor using the following formula: In the formula, Indicates the first The attenuation factor of each grid; Indicates the first Normalized gas disturbance index for each grid; Represented by natural constant An exponential function with base 0.

9. The image stabilization method for UAV inspection according to claim 1, characterized in that, In the step of acquiring the video stream captured by the drone, the video stream is a pre-processed video stream, and the pre-processing operation includes at least one of grayscale processing, adaptive median filtering processing, and histogram equalization processing.

10. An image stabilization system for unmanned aerial vehicle (UAV) inspection, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the image stabilization method for UAV inspection according to any one of claims 1-9.

Images