A power transmission line precise flaw detection system based on an unmanned aerial vehicle stabilizing holder

Abstract

The application relates to the technical field of electric power inspection image processing, and discloses a power transmission line accurate flaw detection system based on a UAV (unmanned aerial vehicle) stabilizing holder, which constructs a local processing area of a to-be-detected fitting and divides key subareas, avoids the interference of background areas, invalid textures and local noises in an original flaw detection image on subsequent processing, introduces attitude prior and structure continuous constraint, performs local image stabilization registration processing on a compression zone subarea, a wire clamp connection zone subarea and a suspected slender abnormal response zone subarea, and improves the alignment stability between multiple key structures, adaptively controls the enhancement intensity based on local structure quality evaluation results and registration reliability, realizes directional enhancement of real boundaries and slender abnormal textures, and suppresses excessive enlargement of low-confidence key subareas and pseudo textures.

Description

Technical Field

[0001] This application relates to the field of power line inspection image processing technology, and in particular to a precision flaw detection system for power transmission lines based on a UAV stabilization gimbal. Background Technology

[0002] Overhead transmission lines operate under high voltage, long spans, and open-air conditions for extended periods. Tension clamps, splicing fittings, and crimping fittings are prone to developing hidden defects such as cracks, strand breakage, incomplete crimping, and deformation after prolonged exposure to mechanical loads, temperature variations, wind vibration, and aging. These defects are often located inside the fittings or obscured by external structures, making them difficult to detect promptly through conventional visual inspections. Traditional manual tower climbing inspections are cumbersome, labor-intensive, time-consuming, and pose significant safety risks. To address the need for condition-based maintenance under uninterrupted power supply conditions, existing technologies have proposed utilizing drones. Equipped with flaw detectors, a measurement and control system, and cameras, live-line flaw detection operations are performed by first placing an imaging plate near the conductor or clamp to be inspected, and then using a drone equipped with X-ray equipment to fly to the target area to take pictures and transmit the images back to the ground platform for analysis. This allows for the acquisition of images of the internal structure of transmission line auxiliary equipment without power outages or disassembly. For this type of operation, existing systems are typically equipped with an electronically controlled gimbal to reduce the impact of flight disturbances on the shooting angle. At the same time, anti-electromagnetic interference design and image preprocessing are combined to improve the engineering feasibility and image availability of high-altitude live-line flaw detection.

[0003] Existing defect detection methods for power transmission lines mainly include ultrasonic imaging, ultrasonic phased array testing, machine vision inspection, infrared inspection, eddy current testing, electric field distribution testing, and X-ray imaging. Among these, ultrasonic imaging and ultrasonic phased array testing are susceptible to corona discharge and electromagnetic interference in live environments, and their application in grid connection is limited. Machine vision inspection mainly targets surface defects and lacks adaptability to internal hidden defects. Infrared inspection is significantly affected by ambient temperature and equipment contamination, making it difficult to reliably reflect internal structural anomalies. Eddy current testing and electric field distribution testing are dependent on conductor material and electromagnetic properties, resulting in poor versatility. While X-ray imaging can visually characterize internal defects such as air gaps, inclusions, displacement, and deformation, it still faces challenges in live-line inspection scenarios using unmanned aerial vehicles (UAVs). The complex environment, difficulties in aerial equipment positioning, high requirements for human-machine collaboration, and electromagnetic interference caused by the direct contact of the imaging plate with high-voltage conductors, along with corona discharge on the conductor surface, spark discharge at the moment of contact with the imaging plate, micro-shaking of the drone during hovering, and slight swaying of the inspected part, can all lead to local blurring, edge distortion, key area offset, and quality fluctuations in the acquired flaw detection images. This makes it difficult to reliably extract effective features from key parts such as tension clamps in subsequent image processing and defect identification, ultimately affecting the accuracy of interpreting internal minor defects and the consistency of detection results. Therefore, how to achieve stable and effective image registration and distortion suppression in live flaw detection images acquired by drone stabilization gimbals has become a key technical problem for improving the reliability of accurate flaw detection in transmission lines. Summary of the Invention

[0004] This application proposes a precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal to address the problems mentioned in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution: a precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal, comprising: The local region construction module is used to acquire the original flaw detection image, construct the local processing area of ​​the hardware to be inspected based on the original flaw detection image, and divide the local processing area of ​​the hardware to be inspected into key sub-regions. The local image stabilization and registration module is used to form an attitude prior based on the residual attitude data corresponding to the original flaw detection image and the preset attitude mapping relationship, and combined with the preset structural continuity constraints, to perform local image stabilization and registration processing on the key sub-region to obtain the registered key sub-region image and registration reliability. An enhancement control module is used to perform local structural quality evaluation processing on the registered key sub-region image to obtain a local structural quality evaluation result, and determine the enhancement intensity based on the local structural quality evaluation result and the registration reliability, and perform local enhancement processing on the registered key sub-region image to obtain an enhanced key sub-region image. The interpretation and fusion module is used to perform filtering and fusion processing based on the local structure quality evaluation results, the registration reliability, and the enhanced key sub-region image to obtain the optimal interpretation image.

[0006] Furthermore, the local region construction module performs a preset relative attenuation normalization process on the original flaw detection image to obtain an attenuation characterization image, and constructs the local processing region of the hardware under inspection based on the preset response function of the hardware under inspection region, according to the phase consistency response, structural tensor consistency response and attenuation gradient response corresponding to the attenuation characterization image.

[0007] Furthermore, the local region construction module, based on preset key sub-region division rules, performs key sub-region division processing on the local processing area of ​​the hardware to be inspected according to the extension direction of the structural skeleton, the boundary turning position, and the local slender abnormal texture response, to obtain the crimping area sub-region, the wire clamp connection area sub-region, and the suspected slender abnormal response area sub-region.

[0008] Furthermore, the local image stabilization and registration module acquires residual attitude data corresponding to the original flaw detection image, and generates the attitude prior based on the residual attitude data according to a preset attitude mapping function; wherein, the residual attitude data includes at least yaw residual, pitch residual, and roll residual.

[0009] Furthermore, the local image stabilization registration module, based on a preset local registration evaluation function, takes the reference key sub-region image as the registration target, and combines the attitude prior and the structural continuity constraint to perform local image stabilization registration processing on each key sub-region. The local registration evaluation function is used to characterize at least the attenuation structure correspondence, edge continuity correspondence, and pose prior consistency between the registered key sub-region image and the reference key sub-region image, and to generate the registered key sub-region image and the registration reliability based on the optimization results corresponding to the local registration evaluation function.

[0010] Furthermore, the enhancement control module extracts key edge continuity index, regional sharpness index, and grayscale distribution consistency index for the registered key sub-region image, and generates the local structure quality evaluation result based on the key edge continuity index, the regional sharpness index, and the grayscale distribution consistency index according to the preset local structure quality evaluation function.

[0011] Furthermore, when generating the local structure quality evaluation result based on the local structure quality evaluation function, the enhancement control module performs preset normalization processing on the key edge continuity index, the regional clarity index, and the grayscale distribution consistency index, and performs weighted combination processing on each normalized index according to preset index weights to generate the local structure quality evaluation result.

[0012] Furthermore, the enhancement control module determines the enhancement intensity based on a preset enhancement intensity mapping function, the local structural quality evaluation result, and the registration reliability, and performs local enhancement processing on the registered key sub-region image based on a preset non-subsampled shear wave enhancement function and the enhancement intensity to obtain the enhanced key sub-region image.

[0013] Furthermore, the interpretation and fusion module determines a reference key sub-region image from the enhanced key sub-region image based on a preset reference key sub-region image determination rule. Then, based on the effective overlap rate of each enhanced key sub-region image relative to the reference key sub-region image, combined with the local structure quality evaluation result and the registration reliability, and based on a preset image filtering and fusion function, it assigns filtering and fusion weights to the enhanced key sub-region image. Finally, it performs filtering and fusion processing on the enhanced key sub-region image according to the filtering and fusion weights to obtain the optimal interpretation image.

[0014] Furthermore, after obtaining the optimal interpretation image, the interpretation fusion module performs anomaly candidate region extraction processing on the optimal interpretation image based on a preset local background estimation function and anomaly evidence function, according to the difference information between the optimal interpretation image and the local background estimation result, to obtain anomaly candidate regions. When the registration reliability of any key sub-region is lower than a preset reliability threshold or the effective overlap rate is lower than a preset overlap rate threshold, the corresponding key sub-region is marked as a region to be reshot.

[0015] The beneficial effects of this invention are as follows: By constructing a local processing area for the hardware to be inspected and dividing it into key sub-regions, interference from background areas, invalid textures, and local noise in the original flaw detection image is avoided for subsequent processing. By introducing attitude priors and structural continuity constraints, local image stabilization and registration processing is performed on the crimping area sub-region, the clamp connection area sub-region, and the suspected slender anomaly response area sub-region, improving the alignment stability between key structures in multiple frames. By adaptively controlling the enhancement intensity based on the local structural quality evaluation results and registration reliability, directional enhancement of real boundaries and slender anomaly textures is achieved, suppressing the excessive amplification of low-confidence key sub-regions and pseudo-textures. By combining the local structural quality evaluation results, registration reliability, and effective overlap rate to perform filtering and fusion processing on the enhanced key sub-region images, the optimal interpretation image with higher stability, more complete coverage, and better suitability for interpretation can be generated. Furthermore, candidate anomaly regions and regions to be re-photographed are output, thereby improving the accuracy of identifying subtle defects inside transmission line hardware, the consistency of interpretation, and the efficiency of flaw detection operations. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort: Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the local image stabilization and registration module of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example

[0018] like Figure 1 and Figure 2 As shown, this invention discloses a precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal, comprising: The local region construction module is used to acquire the original flaw detection image and construct the local processing area of ​​the hardware to be inspected based on the original flaw detection image, and divide the key sub-regions within the local processing area of ​​the hardware to be inspected.

[0019] In this embodiment, the local region construction module is used to acquire the original flaw detection image and construct a local processing area for the hardware to be inspected based on the original flaw detection image. Key sub-regions are then divided within the local processing area of ​​the hardware to be inspected. Here, the original flaw detection image refers to the original imaging result received by the image processing platform after the UAV, equipped with an X-ray pulse machine, irradiates an imaging plate pre-mounted on the side of the hardware to be tested. Since this original flaw detection image not only includes the hardware to be inspected but may also include the background area of ​​the imaging plate, the area adjacent to the conductor structure, and low-frequency non-uniform areas formed by the influence of the charged environment, directly performing subsequent local image stabilization and registration processing on the entire original flaw detection image would be problematic. In the process of quality evaluation, local enhancement, and screening fusion, a large number of background pixels that are not directly related to defect identification will enter the subsequent processing chain. This can easily cause the fine structural boundaries and weak anomalous textures to be covered by background grayscale fluctuations, thereby reducing the targeting and stability of subsequent processing. Therefore, this implementation does not follow the processing method of uniformly enhancing, detecting, or segmenting the entire image in the prior art. Instead, it first constructs the local processing area of ​​the hardware to be inspected from the original flaw detection image, and then further divides the key sub-regions within the local processing area of ​​the hardware to be inspected, so that the subsequent modules always revolve around the crimping area sub-region, the wire clamp connection area sub-region, and the suspected slender anomalous response area sub-region.

[0020] Specifically, the local region construction module first performs a preset relative attenuation normalization process on the original flaw detection image to obtain an attenuation characterization image. The reason for setting this step is that, in addition to reflecting the attenuation effect of the internal structure of the metal under inspection on the radiation, the original gray value of the X-ray flaw detection image is also affected by the detector sensitivity distribution, exposure energy fluctuation, inconsistent local response of the imaging plate, and non-uniform low-frequency background.

[0021] If the original grayscale is directly used to construct the local processing area of ​​the hardware to be inspected, the grayscale comparability of the same type of hardware in different batches of original flaw detection images is poor, which can easily lead to unstable region construction results. Therefore, this implementation converts the original grayscale into an attenuation characterization quantity that is more suitable for comparing local structural differences. In specific implementation, the ratio of the original grayscale value of each pixel to the corresponding empty field reference grayscale value can be calculated, and then the ratio can be negatively logarithmically mapped to obtain the attenuation characterization image. The empty field reference grayscale value here can be derived from the empty field imaging results of the same batch without target occlusion, or it can be derived from the pre-calibrated empty field reference image.

[0022] To avoid numerical instability caused by excessively small original grayscale values ​​or empty field reference grayscale values, a preset minimum grayscale lower limit is introduced during calculation. Preferably, after normalizing the original grayscale values ​​and empty field reference grayscale values ​​to the range of 0 to 1, the minimum grayscale lower limit is set to 0.001 to 0.01, more preferably 0.005. When the minimum grayscale lower limit is less than 0.001, the detector dark current noise and low-count random fluctuations will be abnormally amplified after negative logarithmic mapping. When the minimum grayscale lower limit is greater than 0.01, the weak attenuation difference near the boundary of the metal fitting under test will be compressed, which is not conducive to the subsequent construction of the local processing area of ​​the metal fitting under test. It should be noted that the attenuation characterization image obtained in this embodiment is used to enhance the relative comparability between local structures and is not used for absolute thickness inversion.

[0023] After obtaining the attenuation characterization image, the local region construction module constructs the local processing region of the hardware under test based on the preset response function of the hardware under test region and the phase consistency response, structural tensor consistency response and attenuation gradient response corresponding to the attenuation characterization image. Although the edge detection, Gaussian filtering and simple threshold segmentation in the existing technology can separate high-contrast regions to a certain extent, for the tension clamps, crimping hardware and splicing hardware involved, relying solely on a single gradient feature can easily misidentify high-frequency noise, scattering pseudo-texture or imaging plate edge as the hardware under test region. Relying solely on a single grayscale threshold is also easily affected by differences in hardware material, thickness and exposure.

[0024] Therefore, this implementation does not use a single feature to construct the local processing area of ​​the hardware to be inspected. Instead, it couples the phase consistency response, the structural tensor consistency response, and the attenuation gradient response so that the local processing area of ​​the hardware to be inspected simultaneously satisfies three criteria: edge stability, structural continuity, and significant attenuation change.

[0025] Among them, the phase consistency response is used to characterize the phase alignment degree of the local structure at multiple scales. Since the contour edge, crimping area boundary and connection transition boundary of the hardware under test usually have high consistency in the multi-scale filtering response, while random noise and discrete scattering pseudo-texture often lack stable phase relationship at multiple scales, using the phase consistency response as part of the response function of the hardware under test area is beneficial to improve the noise resistance of the local processing area construction of the hardware under test.

[0026] The structural tensor consistency response is used to characterize the stability of the principal direction of the local gradient field. The hardware under inspection, especially the crimping area and clamp connection area, typically exhibits a continuous metal structure, and its gradient field principal direction shows strong consistency within its local neighborhood. Therefore, the structural tensor consistency response can characterize the degree of local structural continuity. The decay gradient response is used to characterize the intensity of boundary changes in the decay representation image, thereby supplementing the grayscale transition information between the hardware under inspection and the adjacent background. To enable the three types of responses to participate in the calculation of the response function of the hardware under inspection region, the local region construction module first normalizes the phase consistency response, structural tensor consistency response, and decay gradient response respectively, and then... The preset weights are used for weighted combination. The normalization method preferably adopts the minimum-maximum normalization processing within the same frame of the original flaw detection image, so that the three types of responses all fall into the range of 0 to 1. The response function of the hardware under inspection is obtained by weighted summation of the normalized phase consistency response, the normalized structural tensor consistency response, and the normalized decay gradient response, where the sum of the three weights is 1. The weight parameters are set based on the fact that the construction of the local processing area of ​​the hardware under inspection should prioritize the stability of the structural contour, while avoiding noise false enhancement caused by simply relying on local gradients. Therefore, the phase consistency response and the structural tensor consistency response should account for the main proportion, and the decay gradient response is used to supplement the boundary reinforcement effect.

[0027] Preferably, the weights corresponding to the phase consistency response are 0.30 to 0.45, the weights corresponding to the structural tensor consistency response are 0.25 to 0.40, and the weights corresponding to the attenuation gradient response are 0.20 to 0.35. More preferably, the three weights are 0.40, 0.35, and 0.25, respectively. When the weight corresponding to the attenuation gradient response is too large, the response function of the area to be inspected will overly rely on the intensity of local gray-scale changes, easily including scattering noise and the edge of the imaging plate into the local processing area of ​​the area to be inspected. When the weights corresponding to the phase consistency response or the structural tensor consistency response are too low, the stable recognition ability of the continuous metal structure contour will be weakened, resulting in breakage or missed detection in the local processing area of ​​the area to be inspected. Through the above processing, this embodiment transforms the processing idea of ​​whole-image edge detection or whole-image threshold segmentation in the prior art into multi-response coupling construction processing for the target area of ​​the area to be inspected, thereby improving the accuracy and stability of subsequent key sub-region division from the source.

[0028] After obtaining the response value of the area to be inspected, the local region construction module performs region extraction processing on the response map of the area to be inspected to generate the local processing area of ​​the area to be inspected. In specific implementation, adaptive threshold segmentation can be performed on the response map of the area to be inspected to obtain the initial candidate area of ​​the area. Then, hole filling, small connected component removal and boundary smoothing processing can be performed on the initial candidate area of ​​the area to be inspected to obtain a continuous, closed and relatively stable local processing area of ​​the area to be inspected.

[0029] Adaptive threshold segmentation, hole filling, small connected component removal, and boundary smoothing can all be implemented using existing image processing techniques. These techniques, in this embodiment, primarily serve as the foundation for the innovative solution and do not constitute the core innovation of this module. To ensure that the local processing area of ​​the inspected hardware fully covers the hardware without excessively introducing the surrounding background, the ratio of the initial candidate hardware area to the entire original flaw detection image area is preferably controlled between 0.08 and 0.40, more preferably between 0.12 and 0.30. When this ratio is below 0.08, This usually means that the local processing area of ​​the hardware under inspection is too narrow, which may easily miss the complete boundaries of the crimping area and the clamp connection area. When the ratio is higher than 0.40, it means that the local processing area of ​​the hardware under inspection contains too much background, which is not conducive to the subsequent local image stabilization and registration module to perform highly targeted local image stabilization and registration processing on key sub-regions. The value of the area ratio here is based on the common imaging scale distribution of tension clamps, crimping hardware and splicing hardware in airborne X-ray flaw detection images, taking into account the imaging differences of transmission lines of different voltage levels and different types of hardware.

[0030] After constructing the local processing area of ​​the hardware to be inspected, the local area construction module continues to perform key sub-region division processing within the local processing area of ​​the hardware to be inspected, resulting in the crimping area sub-region, the wire clamp connection area sub-region, and the suspected slender abnormal response area sub-region. The reason why it does not stop at obtaining the local processing area of ​​the hardware to be inspected, but continues to divide the key sub-regions, is that the subsequent local image stabilization and registration module does not perform the same local image stabilization and registration processing on the entire local processing area of ​​the hardware to be inspected, and the enhancement control module does not apply a uniform enhancement intensity to the entire local processing area of ​​the hardware to be inspected. Instead, it needs to establish processing objects for the crimping area sub-region, the wire clamp connection area sub-region, and the suspected slender abnormal response area sub-region. If the key sub-region division is not completed in the local area construction module first, the subsequent steps will be difficult to form a progressive processing chain around the key local structure, and will eventually degenerate into coarse-grained whole image processing.

[0031] In specific implementation, the local region construction module performs key sub-region division processing on the local processing area of ​​the hardware to be inspected based on preset key sub-region division rules, according to the extension direction of the structural skeleton, the boundary turning position, and the local slender abnormal texture response. First, the local processing area of ​​the hardware to be inspected is subjected to skeleton refinement processing to obtain the structural skeleton. Skeleton refinement processing itself is an existing image analysis method used to convert the local processing area of ​​the hardware to be inspected into a refined skeleton structure suitable for describing the overall structural extension relationship. Then, the local main extension direction and local width change are calculated along the structural skeleton. Under normal circumstances, the crimping area sub-region corresponds to the main segment on the structural skeleton that is relatively long, has a relatively gentle width change, and has a relatively continuous attenuation characteristic; the clamp connection area sub-region corresponds to the connection transition segment at the bifurcation point, end connection point, or boundary curvature of the structural skeleton; the suspected slender abnormal response area sub-region corresponds to the slender response band that is distributed along the main direction or near the main direction of the hardware to be inspected, has a length that is significantly greater than its width, and has an abnormal attenuation difference from the surrounding homogeneous area.

[0032] To improve the stability of the segmentation, this embodiment further introduces a boundary turning angle threshold, a length-to-width ratio threshold, and a direction deviation threshold. The boundary turning angle threshold is used to characterize the degree of directional change of the contour of the local processing area of ​​the hardware under inspection within the neighborhood of a certain point. In one embodiment, when the boundary turning angle is greater than 20 to 60 degrees, the corresponding neighborhood can be preferentially marked as a candidate area for the wire clamp connection area sub-region. Preferably, the boundary turning angle threshold is 35 degrees. If the threshold is too small, a large number of slight contour undulations will be mistakenly included in the wire clamp connection area sub-region. If the threshold is too large, it is easy to miss the real connection transition position. The length-to-width ratio threshold and the direction deviation threshold are used together to characterize the local thin and elongated abnormal texture response. When the length-to-width ratio of a local connected response band is greater than 3 to 8, and its main direction deviates from the main extension direction of the structural skeleton by no more than 25 to 40 degrees, it can be marked as a candidate region for a suspected slender anomaly response area. Preferably, the length-to-width ratio threshold is 5 and the direction deviation threshold is 30. The reason for adopting this joint constraint is that crack-like, stray, and slit-like anomalies usually appear as relatively slender abnormal attenuation bands in X-ray flaw detection images, rather than near-circular or large blocky anomaly regions. Therefore, the joint constraint of length-to-width ratio and direction consistency can effectively suppress the interference of random noise points and blocky pseudo-textures on the division of suspected slender anomaly response area sub-regions.

[0033] In this embodiment, the division results of the crimping area sub-region, the clamp connection area sub-region, and the suspected slender anomaly response area sub-region are not isolated from each other, but serve the collaborative processing of subsequent modules. Specifically, the crimping area sub-region will serve as the main structural registration object of the subsequent local image stabilization registration module, ensuring high consistency of the crimping boundary area among multiple frames of original flaw detection images; the clamp connection area sub-region will provide structural transitions and connection constraints for the subsequent local image stabilization registration module, preventing the subsequent local image stabilization registration process from optimizing only around the main long strip structure while ignoring the connection parts; the suspected slender anomaly response area sub-region provides the key enhancement object for the subsequent enhancement control module, enabling the enhanced key sub-region image to prioritize the enhancement of crack-like, stray, and fine slit-like anomalies. In other words, the local region construction module in this embodiment does not simply complete simple region clipping, but provides a unified, stable, and traceable local structural object for the subsequent local image stabilization registration module, enhancement control module, and interpretation fusion module through the preprocessing of the local processing region construction and key sub-region division of the inspected hardware.

[0034] The local image stabilization and registration module is used to form an attitude prior based on the residual attitude data corresponding to the original flaw detection image and the preset attitude mapping relationship, and combined with the preset structural continuity constraints, to perform local image stabilization and registration processing on the key sub-region to obtain the registered key sub-region image and registration reliability.

[0035] In this embodiment, the key sub-regions are the crimping area sub-region, the clamp connection area sub-region, and the suspected slender abnormal response area sub-region, which are divided by the local region construction module within the local processing area of ​​the hardware to be inspected. The reason why this module does not perform global registration on the entire original flaw detection image, but instead performs local image stabilization registration processing around the key sub-regions, is that the entire original flaw detection image still contains the imaging plate background area, local low-texture areas, and peripheral structure areas that are not directly related to defect interpretation. If the entire image is directly used as the registration object, background grayscale fluctuations, imaging plate edges, and invalid textures are likely to participate in the matching, thereby interfering with the true alignment of key structures such as crimping boundaries, connection transition boundaries, and slender abnormal textures. By using the key sub-regions output by the local region construction module as the only registration object, the local image stabilization registration processing can always revolve around the local structures in the hardware to be inspected that are directly related to subsequent interpretation.

[0036] Specifically, the local image stabilization and registration module first acquires the residual attitude data corresponding to the original flaw detection image, and generates attitude priors based on the preset attitude mapping function. The residual attitude data here includes at least yaw residuals, pitch residuals, and roll residuals, which are used to characterize the attitude deviations that remain after the active vibration reduction of the stabilization gimbal. The reason for setting this step is that the residual attitude data reflects the actual degree of deviation of the flaw detection imaging component from the ideal imaging attitude, and can provide prior constraints for the local search center offset information and initial rotation estimation information of the key sub-region in the image plane. If the attitude prior is not introduced and blind local registration is performed only based on the image content, then in the case of local texture repetition, edge approximation, or low contrast in the crimping area sub-region and the clamp connection area sub-region, it is easy to have problems such as incorrect matching or convergence to local extrema.

[0037] It should be noted that this implementation does not directly stitch the residual attitude data with the grayscale of the original flaw detection image. Instead, it converts the residual attitude data into local search center offset information and initial rotation estimation information through an attitude mapping function, so that the attitude information participates in the local image stabilization and registration process in the form of geometric prior. The reason for this setting is that the residual attitude data corresponds to the imaging geometric changes, rather than a new source of attenuation characteristics. Therefore, using it as a geometric prior to guide the local search is more in line with the image formation mechanism and the actual scenario.

[0038] In one embodiment, the attitude mapping function is obtained through ground calibration. Specifically, during the calibration stage, the imaging plate and the simulated hardware to be inspected are kept relatively fixed, and the stabilization gimbal is controlled to apply known yaw disturbances, pitch disturbances, and roll disturbances within a small angle range. The displacement and rotation changes of the crimping area sub-region, the clamp connection area sub-region, and the suspected slender abnormal response area sub-region in the image plane are recorded, thereby establishing the correspondence between the residual attitude data and the changes in the image plane of the key sub-regions.

[0039] Preferably, when the yaw and pitch residuals are within the range of -5 to +5, a linear attitude mapping function is used to generate local search center offset information; when the yaw or pitch residuals exceed -5 to +5 but do not exceed -10 to +10, a quadratic correction attitude mapping function is used to generate extended search center offset information; when any yaw or pitch residual exceeds -10 to +10, the corresponding key sub-region of the current frame is marked as a low-confidence candidate object, and its participation priority is reduced in subsequent local image stabilization registration processing. The reason for this segmented setting is that within a small angle range, the displacement change of the key sub-region in the image plane usually approximately satisfies a monotonic mapping relationship with the residual attitude, which can stably provide the initial search direction for local image stabilization registration processing; when the attitude deviation further increases, the risks of local occlusion, target out-of-frame, and projection distortion increase significantly. If the object is still regarded as a high-confidence candidate at this time, it will be considered a low-confidence candidate. Inputting incorrectly can actually reduce the accuracy of subsequent registration reliability determination. For roll residuals, in the preferred embodiment, when the roll residual is within the range of -3 to +3, the initial rotation estimation information is directly generated through the attitude mapping function; when the roll residual exceeds -3 to +3 but does not exceed -6 to +6, it is included in the extended rotation search range; when the roll residual exceeds -6 to +6, the corresponding key sub-region is directly marked as a low-confidence candidate object. This is because after the active vibration reduction of the stabilization gimbal, the residual rotation deviation is usually within a small fluctuation range. Within this range, the key sub-region can still be approximated as a local rigid structure. If the rotation deviation is too large, the approximate premise of the local rigid structure will no longer hold stably. Forcibly performing local image stabilization registration processing can easily introduce mismatch. The records in the reference materials regarding the active vibration reduction function of the stabilization gimbal, which can reduce disturbances and provide a stable shooting angle, provide scenario support for the above settings.

[0040] After generating the attitude prior, the local image stabilization and registration module further determines the reference key sub-region image and, in conjunction with structural continuity constraints, performs local image stabilization and registration processing on each key sub-region. The reference key sub-region image is preferably selected from the key sub-regions corresponding to the original flaw detection images in the same batch. The selection criteria include at least local sharpness, exposure saturation ratio, and key sub-region boundary integrity. In one embodiment, the key sub-region image with the highest local sharpness, exposure saturation ratio below 0.03, and key sub-region boundary integrity above 0.85 can be preferentially selected as the reference key sub-region image. The reason for this setting is that if the reference key sub-region image itself has obvious blur, local overexposure, or missing boundaries, all subsequent key sub-regions will perform local image stabilization and registration processing around an unstable object, thereby reducing the quality of the registered key sub-region image. The local sharpness assessment, exposure saturation ratio statistics, and key sub-region boundary integrity calculation can be obtained using existing image analysis methods, and are mainly used as supporting processing methods in this embodiment.

[0041] After determining the reference key sub-region image, the local image stabilization and registration module, based on a preset local registration evaluation function, uses the reference key sub-region image as the registration target and combines attitude priors and structural continuity constraints to perform local image stabilization and registration processing on each key sub-region. Here, the structural continuity constraints are used to characterize the edge continuity and local structural correspondence of the key sub-region. For the crimping area sub-region, the structural continuity constraints are mainly used to characterize the continuous extension relationship between the crimping boundary and the main body contour, preventing the local image stabilization and registration processing from mismatching the crimping boundary to a nearby but discontinuous position based solely on the similarity of local gray levels. For the clamp connection area sub-region, the structural continuity constraints are mainly used to characterize the turning continuity of the connection transition boundary and the correspondence of the connection direction, preventing directional mismatch at the connection part in the local image stabilization and registration processing. For the suspected slender abnormal response area sub-region, the structural continuity constraints are mainly used to characterize the directional stability and connectivity of the slender abnormal texture in the local neighborhood, thereby suppressing the interference of isolated noise points and discrete pseudo-textures on the local image stabilization and registration processing. It can be seen that the structural continuity constraints are not abstract settings, but directly affect the judgment of the structural correspondence of different key sub-regions.

[0042] To achieve the above processing, this embodiment establishes local similarity transformation models for each key sub-region, ensuring that local image stabilization and registration processing includes at least three degrees of freedom: local translation, local rotation, and local scale fine-tuning. The reason for using local similarity transformation models instead of directly employing high-degree-of-freedom non-rigid transformations is that, given that the stabilization gimbal has eliminated most of the large-amplitude disturbances, the key sub-regions of the inspected hardware in adjacent original flaw detection images can generally still be approximated as locally rigid structures. Using local similarity transformation models can both cover residual small-angle rotations and translations and avoid structural deformation artifacts introduced by high-degree-of-freedom models. Preferably, the local scale variation range is limited to 0.95 to 1.05, more preferably to 0.98 to 1.02; local rotation... The search range is limited to -5 to +5, more preferably to -3 to +3; the local translation search radius is limited to 10 to 60 pixels, more preferably 15 to 40 pixels, based on the imaging plate resolution and the typical scale of the key sub-region in the image. The scale variation range is set in this way because the relative relationship between the source plate and the image plate changes only slightly in a single flaw detection task, and the key sub-region usually does not experience large-scale drift. If the scale variation range is set too wide, the background structure may be mistakenly included in the feasible solution; if the range is set too narrow, it may not be able to cover the true correspondence when there are slight changes in imaging geometry. The local rotation search range is matched with the actual variation range of the roll residual, and the search space is established with priority around small-angle rotation, which is beneficial to balance computational stability and registration rationality.

[0043] The local registration evaluation function is used to quantitatively evaluate the matching degree between the current key sub-region and the reference key sub-region image during the local similarity transformation search process. To avoid the problem of repeated texture mismatch caused by relying solely on gray-level similarity in general image registration, the local registration evaluation function in this embodiment characterizes at least three types of relationships simultaneously. The first type of relationship is the attenuation structure correspondence, which characterizes whether the attenuation representation structure between the current key sub-region and the reference key sub-region image is consistent after the candidate transformation. In specific implementation, this relationship can be measured based on the difference in local attenuation representation or the difference in attenuation gradient. The second type of relationship is the edge continuity correspondence, which characterizes the candidate transformation. The first type of relationship is whether there is a continuous and stable directional correspondence between the edge chain of the subsequent key sub-region and the edge chain of the reference key sub-region image. The second type of relationship is attitude prior consistency, which is used to characterize whether the current candidate transformation deviates from the local search center offset information and initial rotation estimation information given by the residual attitude data and attitude mapping function. Preferably, the above three types of relationships are first normalized and then weighted and combined in the local registration evaluation function. The reason for this setting is that the attenuation structure correspondence, edge continuity correspondence and attitude prior consistency have different dimensions and numerical ranges. If they are not normalized first, it is easy for one type of relationship to meaninglessly dominate the final evaluation result when combining.

[0044] In a preferred embodiment, the weight of the attenuation structure correspondence in the local registration evaluation function is 0.35 to 0.50, the weight of the edge continuity correspondence is 0.25 to 0.40, and the weight of the pose prior consistency is 0.15 to 0.30, more preferably 0.45, 0.35, and 0.20 respectively. This setting is based on the fact that the main goal of local image stabilization registration is to ensure consistency between the true structural boundaries and internal attenuation structures in key sub-regions across multiple frames of original flaw detection images; therefore, the attenuation structure correspondence should have the highest weight. The edge continuity correspondence is used to ensure the geometric rationality of the registration result, thus having the second highest weight. Pose prior consistency is mainly used to guide the search direction and avoid deviating too far from the true imaging geometry; therefore, it is a guiding term rather than an absolutely dominant term. When the pose prior consistency weight is too low, local image stabilization registration is more likely to deviate from the true pose constraint in low-texture or repetitive texture regions. When the pose prior consistency weight is too high, it may over-rely on the pose prior when the image content itself is more reliable, thereby inhibiting the alignment of true details.

[0045] In the actual execution of local image stabilization and registration, the local image stabilization and registration module preferably adopts a two-level search strategy of coarse-to-fine. First, based on the local search center offset information and initial rotation estimation information provided by the attitude prior, a coarse search window is established to perform a candidate local similarity transformation search on the current key sub-region, filtering out candidate solutions that obviously do not satisfy the edge continuity correspondence and attitude prior consistency. Then, a refinement optimization is performed in the neighborhood of the optimal candidate solution obtained by the coarse search to obtain the optimal local registration result of the current key sub-region relative to the reference key sub-region image. The coarse search and refinement optimization can be implemented using existing image registration techniques, such as correlation matching optimization, gradient descent optimization, or pyramid resolution search. These processing methods are basic implementation methods in this embodiment, mainly used to support the overall processing chain of this module. Compared with general registration methods, this embodiment further limits the local search space by attitude prior and structural continuity constraints, and unifies the attenuation structure correspondence, edge continuity correspondence, and attitude prior consistency into the same evaluation system through the local registration evaluation function, so that the local image stabilization and registration processing is more in line with the imaging characteristics of the live flaw detection scenario and the structural characteristics of the hardware to be inspected.

[0046] After obtaining the optimal local registration result, the local image stabilization registration module outputs the registered key sub-region image and further generates a registration reliability score. The reason for setting the registration reliability score is that even with the introduction of attitude priors and structural continuity constraints, some key sub-regions may still fail to register stably due to local occlusion, missing boundaries, abnormal exposure, or insufficient overlap. If only the registered key sub-region image is output without distinguishing its reliability level, the subsequent enhancement control module may over-enhance low-reliability key sub-regions, and the interpretation fusion module may incorrectly include low-reliability key sub-regions in the generation process of the optimal interpretation image. Therefore, this implementation method considers the optimization result corresponding to the local registration evaluation function, the effective overlap rate between the current key sub-region and the reference key sub-region image, and whether the local optimization has converged. The registration reliability is jointly determined. Preferably, when the effective overlap rate is not less than 0.60, the local optimization reaches the convergence condition, and the optimal local registration result satisfies the local rotation and local scale constraints, the corresponding key sub-region is marked as a high-confidence key sub-region. When the effective overlap rate is less than 0.60, or the local optimization does not converge, or the optimal local registration result exceeds the local rotation and local scale constraints, the corresponding key sub-region is marked as a low-confidence key sub-region. The effective overlap rate threshold is preferably set to 0.60 to 0.85, more preferably 0.70. The convergence condition of local optimization can be set to the rate of change of the evaluation function for two consecutive iterations being less than 0.001, or the number of iterations reaching 40 to 80, more preferably 60. This setting can take into account both the computational efficiency and result stability of local image-stabilized registration processing.

[0047] In this embodiment, the registered key sub-region image and registration reliability are not isolated outputs of this module, but rather common inputs to the subsequent enhancement control module and interpretation fusion module. Specifically, the registered key sub-region image provides the subsequent enhancement control module with key local structures under unified coordinates, enabling local structure quality evaluation and local enhancement processing to be based on aligned key boundaries and local abnormal textures. The registration reliability is used to constrain the determination of subsequent enhancement intensity and to filter the weight allocation of the fusion processing, preventing key sub-regions with unreliable local image stabilization registration results from being over-enlarged or incorrectly selected as high-quality interpretation objects in subsequent steps.

[0048] The enhancement control module is used to perform local structural quality evaluation processing on the registered key sub-region image, obtain the local structural quality evaluation result, determine the enhancement intensity based on the local structural quality evaluation result and registration reliability, and perform local enhancement processing on the registered key sub-region image to obtain the enhanced key sub-region image.

[0049] Specifically, the enhanced control module first extracts key edge continuity, regional sharpness, and grayscale distribution consistency indices from the registered key sub-region images. Based on a preset local structure quality evaluation function, it generates local structure quality evaluation results according to these three indices. The reason for using these three indices is that, in this scenario, the main factors affecting subsequent anomaly identification are not just sharpness alone, but also whether the key boundaries directly related to the structure of the hardware under inspection are continuous, whether the local structural details are clear, and whether the grayscale within the homogeneous band is stable. If only a single index is relied upon, it is easy to cause bias. For example, when only regional sharpness is used as the evaluation criterion, key sub-regions with strong high-frequency noise may also receive high scores; when only edge strength is used as the evaluation criterion, discrete pseudo-edges may also be mistaken for real structures; when only grayscale stability is used as the evaluation criterion, key sub-regions with rich details but clear local boundaries may be underestimated. Therefore, this implementation adopts a multi-index joint evaluation method to improve the ability of the local structure quality evaluation results to reflect the actual identification state.

[0050] Among them, the key edge continuity index is used to characterize whether the edge chain directly related to the structure of the hardware under inspection in the key sub-region is continuous, complete and stable. For the crimping area sub-region, the key edge continuity index mainly reflects whether the crimping boundary and the main body outline boundary are continuous; for the wire clamp connection area sub-region, the key edge continuity index mainly reflects whether the connection transition boundary and the turning boundary are complete; for the suspected slender abnormal response area sub-region, the key edge continuity index mainly reflects whether the slender abnormal texture edge extends stably along the main direction.

[0051] In practice, edge extraction can be performed on the registered key sub-region image to obtain an edge chain set. Then, based on the coverage relationship between the sum of connected edge chain lengths and the expected boundary length, and combined with the edge direction dispersion, a key edge continuity index is generated. Preferably, the key edge continuity index is jointly determined by a connected edge coverage term and a direction stability term. The connected edge coverage term reflects the edge chain coverage degree, and the direction stability term suppresses directionally disordered discrete pseudo-edges. To prevent extremely long noisy edge chains from causing abnormally high scores, the connected edge coverage term is preferably limited to the range of 0 to 1, and the direction stability term is preferably reduced as the edge direction variance increases. The reason for this setting is that the real hardware boundary usually has a relatively stable directional extension relationship in the local neighborhood, while random noise and scattering pseudo-edges often have large directional dispersion.

[0052] The regional sharpness index is used to characterize the resolvability of local structural details in the registered key sub-region image. Considering that the main objects of judgment include indentation boundaries, connection boundaries, and elongated anomalous textures, these objects usually exhibit structures with strong local gradients and concentrated edge transitions. Therefore, the regional sharpness index can be constructed based on local gradient energy or high-frequency response intensity. In specific implementation, it is preferable to first perform a light smoothing process on the registered key sub-region image, then calculate the local gradient energy on the smoothed image, and normalize it according to the area of ​​the key sub-region to obtain the regional sharpness index. The reason for performing a light smoothing process first is that the original high-frequency noise will be significantly amplified in the gradient domain. If isolated noise points are not suppressed first, the regional sharpness index is prone to misjudging noise as real details. The light smoothing process can be implemented using existing Gaussian smoothing or median smoothing methods. Preferably, the smoothing scale is 0.6 to 1.5, more preferably 1.0. If the smoothing scale is too small, the high-frequency noise suppression is insufficient; if the smoothing scale is too large, it will weaken the real details of indentation boundaries and elongated anomalous textures.

[0053] The grayscale distribution consistency index is used to characterize whether the attenuation representation of the homogeneous band within the key sub-region is stable. After local image registration, if the homogeneous band without obvious abnormal response still shows large grayscale fluctuations, it indicates that the current key sub-region may still have residual mismatch, scattering interference, or local pseudo-texture. In this case, it is not advisable to directly apply a high enhancement intensity. In specific implementation, a homogeneous band without obvious abnormal response can be selected in the registered key sub-region image, and the relationship between the variance and mean of the attenuation representation within the homogeneous band can be statistically analyzed. Based on this, a grayscale distribution consistency index can be generated. Preferably, the grayscale distribution consistency index decreases as the variance within the homogeneous band increases and is limited to the range of 0 to 1. When the effective area of ​​the homogeneous band is insufficient, or when the homogeneous band is greatly affected by local abnormal response, the local median absolute deviation can be used to replace the variance to construct a robust consistency index. The reason for this setting is that the median absolute deviation is not sensitive to a small number of extreme outliers, which can prevent individual abnormal pixels from incorrectly classifying the entire homogeneous band as a low-consistency object.

[0054] After obtaining the key edge continuity index, regional sharpness index, and grayscale distribution consistency index, the enhancement control module further generates local structure quality evaluation results based on the local structure quality evaluation function. To avoid distortion caused by the different dimensions and value ranges of the three types of indicators, it is preferable to first perform normalization processing on the above three types of indicators, and then perform weighted combination according to the preset index weights. The normalization processing preferably adopts the minimum-maximum normalization processing within the same batch of candidate key sub-regions, so that all three types of indicators fall into the range of 0 to 1. The local structure quality evaluation results can be obtained by weighted summation of the normalized key edge continuity index, the normalized regional sharpness index, and the normalized grayscale distribution consistency index, where the sum of the weights of the three indicators is 1.

[0055] Preferably, the weights for the critical edge continuity index are 0.35 to 0.50, the weights for the regional sharpness index are 0.25 to 0.40, and the weights for the grayscale distribution consistency index are 0.15 to 0.30. More preferably, they are 0.45, 0.35, and 0.20, respectively. This setting is based on the fact that the subsequent judgment objects are mainly indented boundaries, connecting boundaries, and thin, elongated abnormal textures, so critical edge continuity should account for the highest proportion. Regional sharpness directly affects the recognizability of local details, so it accounts for the second proportion. Grayscale distribution consistency is mainly used to suppress low-quality objects from being misjudged as high-quality objects, so as a stability correction term, it accounts for the third proportion. When the weight of the grayscale distribution consistency index is too low, key sub-regions with large residual scattering interference are prone to obtaining unreasonable high scores. When its weight is too high, it may suppress key sub-regions with rich real details but slight fluctuations in local grayscale. Through this joint evaluation method, the local structural quality evaluation results can more realistically reflect whether the current key sub-region is suitable for subsequent directional selective local enhancement processing.

[0056] After obtaining the local structural quality evaluation results, the enhancement control module determines the enhancement intensity based on the local structural quality evaluation results and the registration reliability. The local structural quality evaluation results reflect the readable state of the current critical sub-region, while the registration reliability reflects the reliability of the current critical sub-region at the geometric alignment level. If the enhancement intensity is determined solely based on the local structural quality evaluation results, some critical sub-regions with low quality scores due to local mismatches may be mistakenly identified as needing stronger enhancement, thereby amplifying misaligned structures and pseudo-textures. If the enhancement intensity is determined solely based on the registration reliability, it is impossible to distinguish which critical sub-regions have been aligned but still need enhancement.

[0057] Therefore, this embodiment further employs an enhancement intensity mapping function to jointly determine the enhancement intensity based on the local structural quality evaluation results and registration reliability. Preferably, the enhancement intensity is inversely proportional to the local structural quality evaluation results and directly proportional to the registration reliability. That is, the lower the local structural quality evaluation results and the higher the registration reliability, the greater the corresponding enhancement intensity; conversely, the higher the local structural quality evaluation results or the lower the registration reliability, the smaller the corresponding enhancement intensity. The reason for this setting is that it is only meaningful to apply stronger enhancement to low-quality objects under the premise that the geometric alignment of the key sub-region is reliable. If the registration result of the current key sub-region itself is unreliable, the enhancement process should be suppressed to avoid amplifying the mismatch results.

[0058] In one embodiment, the enhancement intensity mapping function can be set with three types of parameters: a lower limit of enhancement intensity, an upper limit of enhancement intensity, and an enhancement adjustment index. Preferably, the lower limit of enhancement intensity is 0.10 to 0.30, the upper limit of enhancement intensity is 0.80 to 1.50, and more preferably 0.20 and 1.20 respectively; the enhancement adjustment index is 0.80 to 2.50, and more preferably 1.50. If the lower limit of enhancement intensity is too low, even if the current key sub-region is geometrically aligned reliably, the local quality may be poor, and sufficient enhancement may not be obtained. If the upper limit of enhancement intensity is too high, overshooting and false edges are easily introduced in slender abnormal texture regions. The enhancement adjustment index is used to control the sensitivity of enhancement intensity to changes in local structural quality evaluation results and registration reliability. When the enhancement adjustment index is too small, the enhancement intensity changes too slowly, which is not conducive to distinguishing different quality objects. When the enhancement adjustment index is too large, the enhancement intensity is too sensitive to small fluctuations in local scores, which can easily lead to unstable enhancement results.

[0059] After determining the enhancement intensity, the enhancement control module further performs direction-selective local enhancement processing on the registered key sub-region image based on the preset non-subsampled shear wave enhancement function, according to the enhancement intensity, to obtain the enhanced key sub-region image. The reason for choosing the non-subsampled shear wave enhancement function instead of directly using global histogram equalization, fixed sharpening, or ordinary high-pass enhancement is that the crimp boundary in the crimp area sub-region, the connection transition boundary in the clamp connection area sub-region, and the slender abnormal texture in the suspected slender abnormal response area sub-region are all image objects with obvious directionality and local structure. Non-subsampled shear wave decomposition has the characteristics of multi-scale, multi-directional and insensitive to translational changes, which is more suitable for highlighting the edges of slender structures and directional abnormal textures, while suppressing irregular high-frequency noise.

[0060] In practice, non-subsampled shear wave decomposition can be performed on the registered key sub-region image to obtain low-frequency components and high-frequency components at multiple scales and in multiple directions. The low-frequency components mainly reflect the basic attenuation background of the key sub-region, while the high-frequency components mainly reflect local structural details such as indentation boundaries, connection boundaries, and slender anomalous textures. Then, based on the enhancement intensity and orientation fidelity coefficient, directional gain is assigned to each high-frequency component in each direction. The processed high-frequency components are then reconstructed with the low-frequency components to obtain the enhanced key sub-region image. Here, the orientation fidelity coefficient is used to characterize the consistency between the high-frequency component in the current direction and the continuity of the key edge and the local directional stability. The higher the orientation fidelity coefficient, the more likely the high-frequency component in that direction corresponds to the real structural edge or real anomalous texture; the lower the orientation fidelity coefficient, the more likely the high-frequency component in that direction comes from disordered noise or discrete pseudo-textures. By introducing the orientation fidelity coefficient in the enhancement process, all high-frequency components can be avoided from being amplified indiscriminately, thereby improving the targeting of the orientation-selective local enhancement processing to the real structural edge.

[0061] In one embodiment, the orientation fidelity coefficient is limited to the range of 0 to 1. Preferably, the directional high-frequency components with an orientation fidelity coefficient greater than 0.65 are prioritized for enhancement, the directional high-frequency components with an orientation fidelity coefficient between 0.35 and 0.65 are moderately enhanced, and the directional high-frequency components with an orientation fidelity coefficient less than 0.35 are only weakly enhanced or not enhanced at all. The reason for this setting is that real indented boundaries and elongated anomalous textures usually have relatively stable directional extension relationships, and their corresponding directional high-frequency components have high consistency with the key edge continuity results. However, random noise and discrete scattering pseudo-textures usually do not have stable directional aggregation characteristics after directional decomposition. Preferably, the upper limit of directional gain is controlled between 1.20 and 2.50, more preferably between 1.50 and 2.00. If the upper limit of directional gain is too high, high-frequency noise and local ringing effects will be amplified simultaneously. If the upper limit of directional gain is too low, the enhancement of real elongated anomalous textures and weak edges will be insufficient. For low-frequency components, it is preferable to only slightly maintain or not enhance them to prevent the overall background grayscale from rising and reducing the contrast of edge and anomalous responses.

[0062] To prevent low-confidence key sub-regions from being erroneously amplified during the enhancement process, this embodiment also stipulates that when the registration reliability is lower than a preset reliability threshold, the corresponding key sub-region will only undergo low-intensity enhancement or the registered key sub-region image will be directly retained as the enhanced key sub-region image. Preferably, the reliability threshold is between 0.50 and 0.75, and more preferably 0.60. The reason for this setting is that if the current key sub-region does not form a sufficiently reliable geometric correspondence in the local image stabilization registration stage, the subsequent direction-selective local enhancement processing should not continue to enhance its internal high-frequency components. Otherwise, it is easy to amplify misaligned edges, local pseudo-textures, and residual noise. By introducing this constraint, the enhancement control module can form an effective linkage with the local image stabilization registration module, so that the enhancement intensity adjustment is based on a reliable local geometry.

[0063] In this embodiment, the enhanced key sub-region image is not the final output of this module, but an important input of the interpretation fusion module. Specifically, the enhanced key sub-region image will participate in the selection and fusion processing of the optimal interpretation image together with the local structure quality evaluation result, registration reliability and effective overlap rate in the interpretation fusion module. Therefore, the enhancement control module does not only perform general image enhancement, but through three consecutive steps of local structure quality evaluation processing, enhancement intensity mapping and direction-selective local enhancement processing, the enhanced key sub-region image retains the real structure boundary and abnormal texture, while suppressing disordered noise and excessive magnification of low-confidence key sub-regions.

[0064] Through the above processing, the enhancement control module can generate enhanced key sub-region images for the crimping area sub-region, the clamp connection area sub-region, and the suspected slender abnormal response area sub-region, and simultaneously output local structural quality evaluation results that can be directly called by the subsequent interpretation and fusion module. This provides key local image inputs that are of controllable quality, directional fidelity, and structurally stable for the subsequent generation of the optimal interpretation image.

[0065] The interpretation and fusion module is used to perform filtering and fusion processing based on the local structure quality evaluation results, registration reliability, and enhanced key sub-region images to obtain the optimal interpretation image.

[0066] Specifically, the interpretation and fusion module first determines a reference key sub-region image from the enhanced key sub-region images based on a preset reference key sub-region image determination rule. The reason for setting this step is that subsequent effective overlap rate calculation, fusion weight allocation, and multi-frame fusion processing all need to be based on a unified reference object. The reference key sub-region image is preferably selected from the same batch of enhanced key sub-region images. The selection criteria include at least the local structure quality evaluation result, registration reliability, and boundary integrity. Preferably, the enhanced key sub-region image with the highest local structure quality evaluation result, registration reliability higher than 0.70, and boundary integrity not lower than 0.85 is selected as the reference key sub-region image. The reason for this setting is that if the reference key sub-region image itself has obvious boundary defects, local mismatch, or strong local noise, then all subsequent enhanced key sub-region images will be filtered and fused around an unstable object, thereby reducing the stability and interpretability of the optimal interpretation image. Here, boundary integrity is used to characterize the degree of coverage of the key edge chain on the expected boundary, which can be achieved by using existing edge chain coverage statistics or contour closure statistics.

[0067] After determining the reference key sub-region image, the interpretation and fusion module further calculates the effective overlap rate of each enhanced key sub-region image relative to the reference key sub-region image. The effective overlap rate is used to characterize the degree of true overlap between the current enhanced key sub-region image and the reference key sub-region image in a unified coordinate system. The reason for setting this index is that if a certain enhanced key sub-region image has a high local structural quality evaluation result, but the effective overlap rate with the reference key sub-region image is low, it means that the current image only covers a local segment, or there are obvious out-of-view, cropping, or local missing parts. In this case, even if the local details are clear, it is not suitable as the main interpretation basis. In specific implementation, the current enhanced key sub-region image and the reference key sub-region image can be mapped to a unified coordinate system first, and then the ratio between the overlapping area of ​​the effective coverage part of the key sub-region and the effective area of ​​the reference key sub-region can be calculated to obtain the effective overlap rate.

[0068] Preferably, the effective overlap rate is limited to the range of 0 to 1. Preferably, when the effective overlap rate is lower than 0.55, the current enhanced key sub-region image is marked as a low-coverage object; when the effective overlap rate is between 0.55 and 0.75, it is marked as a medium-coverage object; when the effective overlap rate is higher than 0.75, it is marked as a high-coverage object. The basis for this setting is that in scenarios where the key sub-region is small in scale and the boundary interpretation requirements are high, if the effective overlap rate is too low, the current image cannot provide enough complete structural information for subsequent fusion; if the effective overlap rate is high, the image can be used as the main fusion object.

[0069] After obtaining the effective overlap rate, the interpretation and fusion module calculates the screening and fusion weights corresponding to each enhanced key sub-region image based on the preset image screening and fusion function, according to the local structure quality evaluation results, registration reliability, and effective overlap rate. The reason for setting this step is that the final interpretation should not only determine whether a frame is retained based on the level of a single indicator, but should also comprehensively consider the local quality state of the current image, the reliability of geometric alignment, and the completeness of coverage of the reference key sub-region image. If the screening and fusion weights are assigned only based on the local structure quality evaluation results, images that are over-enhanced locally but have incomplete coverage may also receive high weights; if the screening and fusion weights are assigned only based on the registration reliability, images that are geometrically aligned but lack sufficient local detail expression may be over-retained; if the screening and fusion weights are assigned only based on the effective overlap rate, images with large coverage but blurred boundaries may dominate the final output.

[0070] Therefore, this embodiment uses an image filtering and fusion function to jointly evaluate the three types of information. In specific implementation, it is preferable to first normalize the local structure quality evaluation results, registration reliability, and effective overlap rate, and then combine them according to preset weights to generate the filtering and fusion weights corresponding to each enhanced key sub-region image. Preferably, the weight corresponding to the local structure quality evaluation results is 0.35 to 0.50, the weight corresponding to the registration reliability is 0.25 to 0.40, and the weight corresponding to the effective overlap rate is 0.20 to 0.35. More preferably, they are 0.40 and 0.50, respectively. The rationale for setting 35 and 0.25 is that the optimal interpretation image first needs to have a good local structure representation, so the local structure quality evaluation result should account for the highest proportion; secondly, it needs to ensure the reliability of the geometric correspondence, so the registration reliability should account for the second proportion; thirdly, it needs to ensure that the coverage is complete enough, so the effective overlap rate, as a coverage correction term, accounts for the third proportion. When the weight of the effective overlap rate is too low, images that are locally clear but not fully covered are prone to receiving unreasonably high weights; when its weight is too high, images with large areas but weak details will dominate the fusion result.

[0071] In one embodiment, the image filtering and fusion function can be configured with two levels of constraints: a filtering threshold and a fusion threshold. Preferably, when the filtering and fusion weight is below 0.20, the corresponding enhanced key sub-region image is directly excluded from the fusion process; when the filtering and fusion weight is between 0.20 and 0.45, it is used as an auxiliary fusion object; when the filtering and fusion weight is greater than 0.45, it is used as the main fusion object. The reason for setting these two levels of constraints is that, on the one hand, it is necessary to remove low-quality images that are obviously unsuitable as a basis for interpretation, and on the other hand, it is not possible to retain only a single main fusion object and completely discard the other images with supplementary value. Otherwise, the significance of multi-frame information fusion will be weakened. Through this hierarchical filtering method, the complementary information between different enhanced key sub-region images can be retained while ensuring the stability of the optimal interpretation image.

[0072] After completing the screening and fusion weight allocation, the interpretation and fusion module performs fusion processing on the enhanced key sub-region images that have passed the screening to obtain the optimal interpretation image of the sub-region. The fusion processing here is not a simple averaging, but a weighted fusion of different enhanced key sub-region images according to their respective screening and fusion weights. In specific implementation, each enhanced key sub-region image can be mapped to the coordinate system of the reference key sub-region image first, and then the corresponding pixels can be weighted and synthesized according to the screening and fusion weights to obtain the initial fused image.

[0073] To prevent local abnormal brightness values ​​in a certain frame from having an excessive impact on the fusion result, it is preferable to perform quantile truncation processing on the abnormal brightness responses in each enhanced key sub-region image before weighted synthesis. The truncation range is preferably set to the upper 1% to upper 3% brightness range, and more preferably to the upper 2% brightness range. The reason for this setting is that after direction-selective local enhancement processing, local pseudo-bright spots or extreme high-frequency responses may be enhanced to abnormal levels. If they are directly involved in linear weighting, it is easy to cause local bright spots or edge overshoot in the fusion result. For the initial fusion image, it is preferable to further perform small-scale consistent smoothing processing to weaken the weak splicing traces that may be formed at the boundary seams of different enhanced key sub-region images. This small-scale consistent smoothing processing can be implemented by existing small-scale bilateral smoothing methods or guided filtering methods, which are supportive post-processing methods.

[0074] Since the interpretation focuses of the crimping area sub-region, the clamp connection area sub-region, and the suspected slender abnormal response area sub-region are different, this implementation method preferably performs screening and fusion processing on the three types of key sub-regions independently to obtain the corresponding optimal interpretation image of the sub-region. Then, the optimal interpretation images of each sub-region are uniformly backfilled into the local processing area of ​​the hardware to be inspected to obtain the optimal interpretation image. The reason for this setting is that the crimping area sub-region focuses more on the crimping boundary and the main body positioning relationship, the clamp connection area sub-region focuses more on the connection transition boundary and connection reliability, and the suspected slender abnormal response area sub-region focuses more on slender crack-like, stray, or fine slit-like abnormal textures. If the three types of key sub-regions are completely mixed and then uniformly fused, the weight characteristics and enhancement targets of different regions will interfere with each other, which will weaken the local optimal interpretation ability. By fusing the sub-regions independently and then uniformly backfilling them, the most interpretable image expression of each type of key sub-region can be preserved at the same time.

[0075] After obtaining the optimal interpretation image, the interpretation fusion module further performs anomaly candidate region extraction processing on the optimal interpretation image based on the preset local background estimation function and anomaly evidence function to obtain anomaly candidate regions. The reason for setting this step is that although the optimal interpretation image can already serve as the basis for manual interpretation and subsequent state assessment, if anomaly candidate regions can be explicitly extracted from the optimal interpretation image, the burden on operators to search for anomalies in the entire image can be reduced, and the efficiency of anomaly localization can be improved. In specific implementation, the local background estimation function is used to generate the local background estimation result of the current key sub-region based on the optimal interpretation image, and the anomaly evidence function is used to generate anomaly evidence response based on the grayscale difference, edge amplitude, and direction consistency between the optimal interpretation image and the local background estimation result.

[0076] Preferably, the local background estimation function obtains the background estimation result based on local window midpoint statistics or local guided smoothing; the anomaly evidence function comprehensively considers the gray-level difference, edge amplitude, and directional consistency between the optimal interpretation image and the local background estimation result. The reason for this setting is that real abnormal regions usually not only show local gray-level deviations, but are also often accompanied by abrupt edge changes and directional structural changes. Relying solely on gray-level differences can easily misjudge background fluctuations as abnormalities, and relying solely on edge intensity can easily misjudge normal structural boundaries as abnormalities. Therefore, the anomaly evidence function preferably utilizes both gray-level deviation information and structural change information.

[0077] In one embodiment, the abnormal evidence response output by the abnormal evidence function is preferably limited to the range of 0 to 1. When the abnormal evidence response is higher than 0.60, the corresponding region is marked as a first-level abnormality candidate region; when the abnormal evidence response is between 0.40 and 0.60, the corresponding region is marked as a second-level abnormality candidate region; when the abnormal evidence response is lower than 0.40, it is not output as an abnormality candidate region. The reason for this setting is that different levels of abnormal evidence response can provide different priorities for subsequent manual review or automatic assisted evaluation. For first-level abnormality candidate regions, they can be given priority for key review of crack-like abnormalities, scattered strand-like abnormalities, or pressing abnormalities; for second-level abnormality candidate regions, they can be used as supplementary objects of attention. If the abnormal evidence response threshold is set too low, background fluctuations and weak noise regions are easily included in a large number of abnormality candidate regions; if the threshold is set too high, early weak abnormalities may be missed.

[0078] In addition to extracting abnormal candidate regions, this implementation also stipulates that when the registration reliability of any key sub-region is lower than a preset reliability threshold, or when the effective overlap rate of the key sub-region relative to the reference key sub-region image is lower than a preset overlap rate threshold, the corresponding key sub-region is marked as a region to be re-examined. The reason for setting up regions to be re-examined is that even if the current batch of images has undergone local image stabilization registration processing, direction-selective local enhancement processing, and screening fusion processing, there may still be some key sub-regions that do not meet the stable interpretation conditions due to insufficient coverage, local occlusion, or geometric mismatch. If these regions are not explicitly marked, the operator may mistakenly believe that the current most accurate interpretation is the best one. The optimal interpretation image covers all key structures, thus ignoring the need for re-enhancing. Preferably, the reliability threshold is between 0.50 and 0.75, more preferably 0.60; the overlap rate threshold is between 0.55 and 0.80, more preferably 0.65. The reason for this setting is that if the threshold is too low, some low-confidence key sub-regions that should be re-enhanced will be allowed; if the threshold is too high, key sub-regions that are slightly cropped but still interpretable may be unnecessarily identified as areas to be re-enhanced. By setting areas to be re-enhanced, the interpretation fusion module not only outputs the optimal interpretation image and abnormal candidate regions, but also provides feedback to the front-end acquisition process on whether the current batch of data is sufficient, thereby enhancing the closed-loop nature of the entire processing chain.

[0079] In this embodiment, the optimal interpretation image, the anomaly candidate region, and the region to be re-examined together constitute the complete output of the interpretation fusion module. The optimal interpretation image is used to provide operators with a stable and clear representation of key local structures. The anomaly candidate region is used to highlight locations where there may be pressing anomalies, crack-like anomalies, or strand-like anomalies. The region to be re-examined is used to identify key regions where the current data is still insufficient to support stable interpretation. Through the above processing, the interpretation fusion module does not simply overlay the enhanced key sub-region images. Instead, it uses six consecutive steps—referencing key sub-region image determination, effective overlap rate calculation, screening and fusion weight allocation, weighted fusion, anomaly candidate region extraction, and region to be re-examined marking—to ensure that the optimal interpretation image generation process simultaneously considers local quality status, geometric reliability, regional coverage integrity, and anomaly representation ability. This provides stable, intuitive, and targeted image evidence for subsequent hardware defect status assessment.

[0080] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal, characterized in that, include: The local region construction module is used to acquire the original flaw detection image, construct the local processing area of ​​the hardware to be inspected based on the original flaw detection image, and divide the local processing area of ​​the hardware to be inspected into key sub-regions. The local image stabilization and registration module is used to form an attitude prior based on the residual attitude data corresponding to the original flaw detection image and the preset attitude mapping relationship, and combined with the preset structural continuity constraints, to perform local image stabilization and registration processing on the key sub-region to obtain the registered key sub-region image and registration reliability. An enhancement control module is used to perform local structural quality evaluation processing on the registered key sub-region image to obtain a local structural quality evaluation result, and determine the enhancement intensity based on the local structural quality evaluation result and the registration reliability, and perform local enhancement processing on the registered key sub-region image to obtain an enhanced key sub-region image. The interpretation and fusion module is used to perform filtering and fusion processing based on the local structure quality evaluation results, the registration reliability, and the enhanced key sub-region image to obtain the optimal interpretation image.

2. The precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 1, characterized in that, The local region construction module performs a preset relative attenuation normalization process on the original flaw detection image to obtain an attenuation characterization image, and constructs the local processing region of the hardware under inspection based on the preset response function of the hardware under inspection region, according to the phase consistency response, structural tensor consistency response and attenuation gradient response corresponding to the attenuation characterization image.

3. The precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 2, characterized in that, The local region construction module performs key sub-region division processing on the local processing area of ​​the hardware to be inspected based on preset key sub-region division rules, according to the extension direction of the structural skeleton, the boundary turning position, and the local slender abnormal texture response, to obtain the crimping area sub-region, the wire clamp connection area sub-region, and the suspected slender abnormal response area sub-region.

4. The precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 1, characterized in that, The local image stabilization and registration module acquires the residual attitude data corresponding to the original flaw detection image, and generates the attitude prior based on the residual attitude data according to the preset attitude mapping function; wherein, the residual attitude data includes at least yaw residual, pitch residual and roll residual.

5. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 4, characterized in that, The local image stabilization and registration module, based on a preset local registration evaluation function, takes the reference key sub-region image as the registration target, and combines the pose prior and the structural continuity constraint to perform local image stabilization and registration processing on each key sub-region. The local registration evaluation function is used to characterize at least the attenuation structure correspondence, edge continuity correspondence, and pose prior consistency between the registered key sub-region image and the reference key sub-region image, and to generate the registered key sub-region image and the registration reliability based on the optimization results corresponding to the local registration evaluation function.

6. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 1, characterized in that, The enhancement control module extracts key edge continuity index, region sharpness index, and grayscale distribution consistency index from the registered key sub-region image, and generates the local structure quality evaluation result based on the key edge continuity index, the region sharpness index, and the grayscale distribution consistency index according to a preset local structure quality evaluation function.

7. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 6, characterized in that, When generating the local structure quality evaluation result based on the local structure quality evaluation function, the enhancement control module performs preset normalization processing on the key edge continuity index, the regional clarity index, and the grayscale distribution consistency index, and performs weighted combination processing on each index after normalization processing according to preset index weights to generate the local structure quality evaluation result.

8. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 7, characterized in that, The enhancement control module determines the enhancement intensity based on a preset enhancement intensity mapping function, the local structural quality evaluation result, and the registration reliability. Based on a preset non-subsampled shear wave enhancement function, it performs local enhancement processing on the registered key sub-region image according to the enhancement intensity to obtain the enhanced key sub-region image.

9. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 8, characterized in that, The interpretation and fusion module determines a reference key sub-region image from the enhanced key sub-region image based on a preset reference key sub-region image determination rule. Then, based on the effective overlap rate of each enhanced key sub-region image relative to the reference key sub-region image, combined with the local structure quality evaluation result and the registration reliability, it assigns screening and fusion weights to the enhanced key sub-region image according to a preset image screening and fusion function. Finally, it performs screening and fusion processing on the enhanced key sub-region image according to the screening and fusion weights to obtain the optimal interpretation image.

10. A precision flaw detection system for power transmission lines based on an unmanned aerial vehicle (UAV) stabilization gimbal according to claim 9, characterized in that, After obtaining the optimal interpretation image, the interpretation fusion module performs anomaly candidate region extraction processing on the optimal interpretation image based on a preset local background estimation function and anomaly evidence function, according to the difference information between the optimal interpretation image and the local background estimation result, to obtain anomaly candidate regions. When the registration reliability of any key sub-region is lower than a preset reliability threshold or the effective overlap rate is lower than a preset overlap rate threshold, the corresponding key sub-region is marked as a region to be reshot.

Images