Image-based overhead strand breakage detection method and system

Abstract

The application belongs to the technical field of image processing, and particularly relates to an overhead strand broken strand detection method and system based on images, which comprises the following steps: performing multi-scale and multi-direction decomposition on a gray image of an overhead strand to obtain original response components of the gray image in multi-scale and multi-direction; constructing a local harmonic model for representing the direction and frequency characteristics of normal strand texture in a region where a pixel point is located according to the local texture characteristics of the gray image; calculating non-harmonic weights of each original response component according to the local harmonic model and the original response components in multi-scale and multi-direction; performing weighted processing on the original response components in multi-scale and multi-direction according to the non-harmonic weights to construct an abnormal feature map; and locating a broken strand defect according to the abnormal feature map. The application solves the problem of background submerging a target in a dynamic blurred image, and improves the accuracy of overhead strand broken strand defect detection.

Description

Technical Field

[0001] This invention relates to the field of image processing technology. More specifically, this invention relates to an image-based method and system for detecting broken strands in overhead stranded wires. Background Technology

[0002] Overhead stranded conductors bear the main mechanical load in power transmission lines, and their structural integrity directly affects the safe operation of the power grid. With the development of UAV inspection technology, acquiring conductor images using high-resolution cameras has become the mainstream method. However, during flight, the relative motion of the images caused by the vibration of the aircraft and wind-induced swaying results in dynamic blurring in the acquired images, which severely degrades the information of broken strand edges. Traditional edge detection algorithms rely on grayscale gradient intensity, but dynamic blurring significantly weakens high-frequency components, causing gradient response decay, making it difficult to identify minute broken strand defects, resulting in a high rate of missed detection.

[0003] To address this, some studies have introduced phase consistency techniques for feature extraction. This method locates edges by analyzing the phase alignment of multi-scale frequency components and exhibits strong robustness to illumination changes and blurring. However, traditional phase consistency algorithms indiscriminately accumulate responses from all directions and frequencies, lacking the ability to distinguish between normal patterns and abnormal deviations. Normal strands are arranged in a periodic spiral, with concentrated and high-intensity texture response energy. Broken strands, on the other hand, manifest as local burrs or gaps, which are non-periodic disturbances with weak signals. Under dynamic blurring conditions, the frequency domain response of broken strands is further weakened, resulting in a strong periodic background completely masking the weak abnormal signal in the final generated feature map. This creates a phenomenon where the background overwhelms the target, making it impossible to effectively separate the true defect and leading to missed or false detections of broken strand defects in overhead stranded cables.

[0004] Therefore, there is an urgent need for an image-based method and system for detecting broken strands in overhead stranded wires. Summary of the Invention

[0005] To address the technical problem that the aforementioned phase consistency algorithm cannot distinguish between normal textures and abnormal defects, resulting in a strong periodic background overshadowing weak breakage defect signals, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides an image-based method for detecting broken strands in overhead stranded wires, comprising:

[0007] A multi-scale, multi-directional decomposition is performed on the grayscale image of the overhead stranded wire to obtain the original response components of the grayscale image in multiple scales and directions. Based on the local texture characteristics of the grayscale image, a local harmonic model is constructed for each pixel in the grayscale image. The local harmonic model is used to characterize the direction and frequency characteristics of the normal strand texture in the region where the pixel is located. Based on the local harmonic model and the original response components in multiple scales and directions, the non-harmonic weights of each original response component are calculated. The non-harmonic weights are used to characterize the degree of deviation between the original response component and the local harmonic model. The original response components in multiple scales and directions are weighted according to the non-harmonic weights to construct an anomaly feature map. Based on the anomaly feature map, strand breakage defects are located.

[0008] This invention accurately describes the frequency and direction characteristics of local variations in the texture of normal strands in an image by constructing a dedicated local harmonic model for each pixel. Based on the non-harmonic weights calculated from the local harmonic model, it adaptively identifies response components that deviate from the normal texture pattern, i.e., potential defect signals. In the subsequent construction of the abnormal feature map, the non-harmonic weights are used to weight the response components, suppressing strong periodic background signals that match the local harmonic model, while selectively enhancing weak non-periodic strand breakage signals that deviate from the local harmonic model. This effectively solves the problem of background obscuring the target, improves the contrast of defect signals, and makes the detection of strand breakage defects in overhead stranded cables more accurate.

[0009] Preferably, the step of performing multi-scale, multi-directional decomposition on the grayscale image of the overhead stranded line to obtain the original response components of the grayscale image at multiple scales and in multiple directions includes: filtering the grayscale image of the overhead stranded line using a multi-scale, multi-directional Log-Gabor filter bank to obtain complex response matrices at each scale and in each direction; and separating the amplitude component and phase component from the complex response matrices at each scale and in each direction as the original response components at each scale and in each direction.

[0010] Preferably, the step of constructing a local harmonic model for pixels in a grayscale image based on the local texture characteristics of the grayscale image includes: constructing a structure tensor based on the gradient information of each pixel in the grayscale image; performing eigenvalue decomposition on the structure tensor to obtain an optimized local texture direction; determining the direction index with the smallest angular difference from the optimized local texture direction among multiple directions; obtaining the amplitude of the pixel at the direction index at multiple scales; obtaining the scale corresponding to the largest amplitude; using the center frequency of the Log-Gabor filter corresponding to that scale as the local texture frequency of the pixel; and combining the optimized local texture direction and the local texture frequency to form a local harmonic model for the pixel.

[0011] This invention constructs a structural tensor based on pixel gradient information and performs eigenvalue decomposition, which can accurately obtain the texture direction information of each pixel. Combined with the optimization processing of texture direction, it can effectively eliminate the direction jitter problem caused by dynamic blur. After determining the direction that best matches the optimized texture direction, by analyzing the response amplitude at different scales in this direction, the frequency corresponding to the scale with the strongest energy is selected as the local texture frequency. This can capture the texture characteristics of overhead stranded wires at different locations, laying the foundation for subsequent differentiation between normal texture and strand breakage defects.

[0012] Preferably, obtaining the optimized local texture direction includes: decomposing the original local texture direction at the pixel into horizontal vector components and vertical vector components; performing Gaussian smoothing on the horizontal and vertical vector components of all pixels respectively to obtain smoothed horizontal and vertical vector components; and combining the smoothed horizontal and vertical vector components to obtain the optimized local texture direction.

[0013] This invention converts the original local texture direction into a vector representation and then performs smoothing, which can effectively solve the periodicity problem of the direction angle and avoid the discontinuity caused by directly smoothing the angle. By performing Gaussian smoothing on the horizontal and vertical vector components respectively, the local direction jitter caused by dynamic blur and noise can be eliminated, so that the texture direction remains continuously changing in space, improving the reliability and continuity of the direction estimation.

[0014] Preferably, the non-harmonic weights satisfy the expression: It should be noted that, For pixels In scale and direction Nonharmonic weights; For the index of the preset direction, For pixels Optimize local texture direction, For direction and optimize local texture direction The minimum angular difference between them; For scale The corresponding center frequency of the Log-Gabor filter; For pixels Local texture frequencies; It is the natural logarithm function; It is a natural exponential function; These are directional bandwidth control parameters; These are frequency bandwidth control parameters.

[0015] This invention achieves selective response to normal textures and anomalous features by measuring the deviation of the current response component from the local harmonic model. When the response component matches the local model, it is given low weight, and when the response component deviates from the local model, it is given high weight, which effectively solves the problem of strong periodic textures drowning out weak strand breakage signals in traditional phase consistency algorithms. By combining deviation measurements in two dimensions, direction and frequency, the algorithm can identify potential strand breakage signals that are inconsistent with normal strand textures.

[0016] Preferably, the anomaly feature map is constructed by weighting the original response components of multiple scales and directions according to the non-harmonic weights, including: weighting the original response components of multiple scales and directions according to the non-harmonic weights to obtain the initial anomaly response value of the pixel; constructing the model confidence of the pixel according to the amplitude of the pixel at all scales along the direction index; taking the product of the model confidence and the initial anomaly response value as the final anomaly response value of the pixel; and constructing the anomaly feature map from the final anomaly response values ​​of all pixels.

[0017] This invention applies non-harmonic weights to the original response components, which can suppress normal texture responses that match the local harmonic model and enhance anomalous feature responses that do not match the model. The introduction of a model confidence mechanism can effectively evaluate the reliability of the local model and avoid false alarms caused by model failure in flat or textureless regions. Multiplying the initial anomalous response value with the model confidence ensures that only anomalous signals appearing in reliable regions of the model are retained, thereby improving the signal-to-noise ratio and reliability of the anomalous feature map.

[0018] Preferably, the initial anomaly response value satisfies the expression:

[0019] ;in, For pixels The initial abnormal response value; For pixels In scale and direction Nonharmonic weights; For pixels In scale and direction The range of the above; Represents pixels In scale and direction Phase on, It is a cosine function. It is a sine function; Indicates the scale quantity. Indicates the number of preset directions; This means that only positive values ​​are considered, and negative values ​​are counted as 0. To prevent extremely small positive numbers with a denominator of zero.

[0020] This invention combines non-harmonic weights, amplitude, and phase consistency to calculate the initial abnormal response value, achieving selective enhancement of abnormal features. The introduction of the phase consistency mechanism ensures the algorithm's robustness to dynamic fuzziness, because phase information is more stable than amplitude information under fuzzy conditions, effectively highlighting the feature responses that deviate from the local harmonic model, and providing high-quality abnormal feature maps for subsequent defect localization.

[0021] Preferably, the model confidence level satisfies the expression: ;in, For pixels Model confidence at the location; For pixels In scale and direction The range of the upper part, for In each preset direction and pixel point The index of the direction with the smallest local texture direction angle difference is optimized. For pixels In direction The index of the scale corresponding to the largest amplitude among all scales; For pixels In scale and direction The range of the upper part, For pixels In direction The sum of the amplitudes of all scales above, Indicates the scale quantity. Indicates the number of preset directions.

[0022] Preferably, locating the broken strand defect based on the abnormal feature map includes: performing Otsu threshold segmentation on the abnormal feature map to obtain a binary image of the defect; performing connected component analysis on the binary image of the defect to obtain multiple connected components; and, in response to a connected component having an area greater than a preset minimum area threshold and less than a preset maximum area threshold, designating the connected component as the broken strand defect region.

[0023] Secondly, the present invention provides an image-based overhead strand breakage detection system, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned image-based overhead strand breakage detection method is implemented.

[0024] By adopting the above technical solution, a computer program for detecting broken strands in overhead stranded wires based on images is generated and stored in a memory so that it can be loaded and executed by a processor. This allows for the creation of a terminal device based on the memory and processor, making it convenient to use.

[0025] The beneficial effects of this invention are as follows: By constructing a local harmonic model, this invention achieves adaptive characterization of the normal texture characteristics of overhead stranded wires; based on this model, a non-harmonic weight field is designed so that phase consistency calculation only responds to non-periodic components that deviate from the normal pattern, fundamentally solving the problem of strong background overwhelming weak strand breakage signals; the generated abnormal feature map has a high signal-to-noise ratio, facilitating subsequent stable segmentation; this method has good robustness to dynamic fuzziness, illumination changes, and conductor deformation, is suitable for complex field inspection environments, can be integrated into UAVs or ground monitoring systems, and has good engineering application prospects. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an image-based method for detecting broken strands in overhead stranded wires according to the present invention.

[0027] Figure 2 This is a schematic grayscale image of an overhead stranded wire;

[0028] Figure 3 It is a visualization that schematically shows the model confidence level;

[0029] Figure 4 This is a schematic illustration of the detection results for strand breakage defects. Detailed Implementation

[0030] 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, not all, of the embodiments of the present invention. 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.

[0031] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0032] This invention discloses an image-based method for detecting broken strands in overhead stranded cables, referring to... Figure 1 This includes steps S1-S5:

[0033] S1. Acquire grayscale images of overhead stranded lines.

[0034] Specifically, using a drone inspection system or a ground-based telephoto camera, inspection videos of the overhead stranded cable area are collected. The videos are decoded into continuous image frames. Any frame to be inspected is used as the original color image. The original color image is then converted to grayscale to obtain a grayscale image of the overhead stranded cable. For example, Figure 2 This is a grayscale image of an overhead stranded cable.

[0035] Preferably, due to camera shake, camera body displacement, and cable vibration caused by strong winds during overhead stranded line inspection, the original acquired images inevitably suffer from dynamic blur and may contain isolated noise points introduced by sensors or signal transmission, such as salt-and-pepper noise. These interference factors can affect subsequent gradient-based local orientation field estimation, leading to inaccurate feature extraction. The median filter, a nonlinear filter, can effectively remove isolated salt-and-pepper noise while having minimal impact on the edge blurring characteristics introduced by dynamic blur. Therefore, this invention performs median filtering on the grayscale images of overhead stranded lines.

[0036] S2. Perform multi-scale and multi-directional decomposition on the grayscale image of the overhead stranded line to obtain the original response components of the grayscale image in multiple scales and directions.

[0037] It should be noted that the phase consistency algorithm does not rely on the blurred and corrupted image gradient information. Instead, it utilizes robust phase information to ensure feature extraction capabilities on dynamically blurred images, and can be used to extract edge and line features from grayscale images of overhead stranded lines. However, since traditional phase consistency algorithms indiscriminately accumulate all responses, strong periodic strand textures can overwhelm weak strand breakage signals. Therefore, this invention only decomposes the grayscale image of the overhead stranded line into multi-scale, multi-directional original response components without accumulation, preserving all original information for subsequent model-based selective weighting.

[0038] Specifically, a Log-Gabor filter bank is used to filter the grayscale image of the overhead stranded line. The Log-Gabor filter bank includes... Log-Gabor filters, where Indicates the scale quantity. This indicates the number of preset directions. The preset directions are determined by... Discretize the angle range as The angles are obtained by equal intervals.

[0039] It should be noted that each Log-Gabor filter has a Gaussian response on the logarithmic frequency scale, has no DC component, and has controllable bandwidth, making it suitable for extracting local frequency information from images. Log-Gabor filters at different scales have different center frequencies; smaller scales correspond to higher frequencies, capable of capturing finer texture details in overhead wires, such as the burrs on broken strands, while larger scales correspond to lower frequencies, capable of capturing coarser texture structures in overhead wires, such as the overall outline of the strands. In this embodiment, the number of scales... The setting is 5 to ensure coverage of the overhead stranded wire's texture features, from fine to coarse. In this embodiment, the number of preset directions is set to 8, allowing the Log-Gabor filter to cover 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315° directions, ensuring the capture of the overhead stranded wire's texture features in each direction. The minimum wavelength of the Log-Gabor filter determines the highest frequency resolution of the image features, and the minimum texture detail of the overhead stranded wire is typically no less than 3 pixels; therefore, the minimum wavelength is set to 3 pixels. The inter-scale scaling factor of the Log-Gabor filter is set to 1.8 to ensure the continuity of frequency coverage across scales and avoid information loss. The frequency bandwidth of each Log-Gabor filter is set to 0.75 to balance the filter's frequency selectivity and frequency coverage range. In other embodiments, implementers can set the number of scales, the number of directions, the minimum wavelength of the Log-Gabor filter, the inter-scale scaling factor, and the frequency bandwidth according to the actual implementation situation.

[0040] Furthermore, a Fourier transform is performed on the grayscale image of the overhead stranded line, and the result of the Fourier transform is compared with... Multiplying the Log-Gabor filters together and then performing an inverse Fourier transform yields the result. A complex response matrix.

[0041] The amplitude components are separated from each complex response matrix. and phase components ,in, Represents pixels In scale and direction The amplitude reflects the pixel. In scale and direction Energy intensity on Represents pixels In scale and direction The phase on the surface reflects the pixel. The standard for treatment is And the direction is The spatial location relationship of local frequency features extracted by the Log-Gabor filter.

[0042] It should be noted that since the amplitude increases with the texture energy and the phase changes with the feature location, the amplitude and phase components can jointly describe the local features of an image. In image processing, phase information encodes the precise location information of features. When the phases of multiple frequency components are aligned at a pixel, it indicates the presence of a significant feature, such as an edge or line. Under dynamic blur conditions, amplitude information degrades significantly, but phase information remains relatively stable, providing a foundation for subsequent segment breakage detection.

[0043] S3. Based on the local texture characteristics of the grayscale image, construct a local harmonic model for the pixels in the grayscale image. The local harmonic model is used to characterize the direction and frequency characteristics of the normal line texture in the region where the pixel is located.

[0044] It should be noted that, because overhead stranded cables often exhibit sag or perspective distortion in images, the direction and frequency of their normal strand texture vary locally at different pixel locations. A globally fixed model cannot adapt to these variations, resulting in incomplete suppression of normal texture and the potential suppression of weak strand breakage signals. Therefore, this invention dynamically constructs a dedicated local harmonic model for each pixel and extracts the texture direction and frequency characteristics of the region where that pixel is located through structural tensor analysis and gradient information. Simultaneously, the confidence level of the model is evaluated, thereby accurately suppressing the normal texture background and effectively extracting weak strand breakage signals.

[0045] Specifically, the horizontal and vertical gradients of each pixel in the grayscale image of the overhead stranded line are obtained. A structure tensor is constructed based on the horizontal and vertical gradients. The structure tensor is Gaussian smoothed. The smoothed structure tensor is decomposed using an eigenvalue decomposition algorithm to obtain the eigenvector corresponding to the largest eigenvalue. The eigenvector is converted into a direction angle and used as the original local texture direction at the pixel.

[0046] Furthermore, for any pixel , to pixel Original local texture direction Decomposed into two-dimensional vector components ,in, For pixels The horizontal vector component of the original local texture direction, , For pixels The vertical vector component of the original local texture direction. , It is a cosine function. It is a sine function.

[0047] Gaussian smoothing is applied to the horizontal and vertical vector components of the original local texture direction for all pixels. The smoothed horizontal and vertical vector components are then combined to obtain the optimized local texture direction for each pixel. ,in, For pixels Optimize local texture direction, For pixels The smoothing value of the vertical vector component of the original local texture direction. For pixels The smoothing value of the horizontal vector component of the original local texture direction. It is the arctangent function in the four quadrants.

[0048] It should be noted that, since the twisted wire texture direction should change smoothly and continuously, the operation of obtaining the optimized local texture direction in this invention effectively eliminates local jitter, improves the reliability of direction estimation, and makes the change of the optimized local texture direction more continuous, thus ensuring the stability of the subsequent model.

[0049] Furthermore, for each pixel, in Among the preset directions, determine the index of the direction with the smallest angle difference from the optimized local texture direction of the pixel. Get pixel points Direction at each scale The amplitude is calculated, and the scale corresponding to the largest amplitude is obtained. The center frequency of the Log-Gabor filter corresponding to this scale is taken as the local texture frequency of the pixel. Indicates the scale quantity. Indicates the number of preset directions.

[0050] The local harmonic model of each pixel is formed by optimizing the local texture direction and local texture frequency of each pixel.

[0051] Furthermore, to measure the reliability of local texture frequency estimation, the model confidence level for each pixel is obtained:

[0052] ;

[0053] in, For pixels The model confidence score at a given location reflects the reliability of the local texture frequency estimation, and its value ranges from [value range missing]. ; For pixels In scale and direction The range of the upper part, for In each preset direction and pixel point The index of the direction with the smallest local texture direction angle difference is optimized. For pixels In direction The index of the scale corresponding to the largest amplitude among all scales; For pixels In scale and direction The range of the upper part, For pixels In direction The sum of the amplitudes of all scales above, Indicates the scale quantity. Indicates the number of preset directions. When near When, it indicates that energy is highly concentrated on one scale, pixel. The estimation of local texture frequencies is very reliable; when near When, it means that energy is uniformly distributed across all scales, pixel. The estimation of local texture frequencies is unreliable.

[0054] For example, Figure 3 A visualization of the model confidence. Figure 3 The brighter the color, the higher the model confidence; the darker the color, the lower the model confidence. It can be seen that the model confidence is high in the normal stranded area and low in the broken strand area.

[0055] S4. Based on the local harmonic model and the original response components of multiple scales and directions, calculate the non-harmonic weights of each original response component. The non-harmonic weights are used to characterize the degree of deviation between the original response components and the local harmonic model. The original response components of multiple scales and directions are weighted according to the non-harmonic weights to construct an anomaly feature map.

[0056] It should be noted that, since the traditional phase consistency algorithm accumulates all response components indiscriminately, the strong periodic strand texture overwhelms the weak strand breakage signal. Therefore, this invention uses a local harmonic model to obtain the non-harmonic weights of pixels and assigns high weights only to response components that deviate from the normal texture model.

[0057] Specifically, the non-harmonic weights of pixels satisfy the expression:

[0058] ;

[0059] It should be noted that, For pixels In scale and direction Nonharmonic weights; For the index of the preset direction, For pixels Optimize local texture direction, For direction and optimize local texture direction The minimum angular difference between them; For scale The corresponding center frequency of the Log-Gabor filter; For pixels Local texture frequencies; It is the natural logarithm function. Used to measure center frequency With local texture frequency Since the perception characteristics of frequency follow a logarithmic law, this invention uses the natural logarithm function to convert the relative proportional relationship of frequencies into a linear difference, so that the relative changes in the high-frequency and low-frequency regions can be balanced, avoiding the excessive influence of the absolute frequency difference on the high-frequency region. It is a natural exponential function; This is a directional bandwidth control parameter used to adjust the sensitivity to directional deviation. In this embodiment, Set as This value adapts to the typical directional variation range of the texture of overhead stranded cables; This is a frequency bandwidth control parameter used to adjust the sensitivity to frequency deviations. In this embodiment, Set as This value adapts to the typical frequency variation range of overhead stranded wire texture. In other embodiments, implementers can set it according to the actual implementation situation. as well as The smaller the parameter setting, the more sensitive the model is to deviations, and the more suitable it is for regular textures. The larger the parameter setting, the higher the model tolerance, and the more suitable it is for textures with greater variations.

[0060] It should be noted that this invention obtains the non-harmonic weights through a Gaussian function, assigning low weights when the response component matches the local harmonic model and high weights when the response component deviates from the local harmonic model. (Direction deviation portion) Measured the direction With local texture direction The degree of deviation, frequency deviation part Measured the scale Center frequency and local texture frequency The degree of deviation, This represents the absolute deviation of the frequency ratio on a logarithmic scale, when When it equals 1, that is, when the frequency is perfectly matched, the value is 0; when When the value is greater than 1 or less than 1, the value is positive, and it increases with... and The ratio increases monotonically away from 1. When the Log-Gabor filter is scaled... direction When the response component at a given point perfectly matches the local harmonic model, the directional difference and frequency ratio are both close to zero, and the exponent term is close to 1, making... Approaching 0; when the Log-Gabor filter is scaled direction When the response components at a certain point deviate significantly from the local harmonic model, the directional difference and frequency ratio increase, and the exponential term approaches zero, leading to... With a value close to 1, this design can suppress normal textures and selectively enhance abnormal features, effectively solving the problem of strong periodic textures overwhelming weak breakage signals in traditional phase consistency algorithms.

[0061] Furthermore, the initial abnormal response value of the pixel is calculated based on the non-harmonic weight of the pixel:

[0062] ;

[0063] in, For pixels The initial abnormal response value; For pixels In scale and direction Nonharmonic weights; For pixels In scale and direction The range of the above; Represents pixels In scale and direction Phase on, It is a cosine function. It is a sine function; Indicates the scale quantity. Indicates the number of preset directions; This means that only positive values ​​are considered, and negative values ​​are counted as 0. To prevent extremely small positive numbers with a denominator of zero, in this embodiment, Set as In other embodiments, implementers may set the appropriate parameters according to the actual implementation situation. .

[0064] It should be noted that this invention achieves suppression of normal textures and selective enhancement of anomalous features through non-harmonic weighting. During the calculation of the initial anomalous response value, High weights are assigned to response components that deviate from the normal texture model. Indicates the magnitude of the response. The phase consistency metric is larger when the phases of multiple frequency components are spatially aligned. The product of the three ensures that only features that do not match the local model but have phase consistency can generate an effective response. This solves the problem of strong periodic textures overwhelming weak breakage signals in traditional phase consistency algorithms. Under dynamic fuzzy conditions, although amplitude information is severely degraded, phase information remains relatively stable. By combining non-harmonic weights and phase consistency metrics, weak breakage signals that are overwhelmed by normal thread textures can be effectively extracted, providing high-quality abnormal feature maps for subsequent defect localization.

[0065] Furthermore, the initial anomaly response values ​​are corrected using the model confidence scores of each pixel to obtain the final anomaly response values ​​for each pixel. These final anomaly response values ​​are then used to construct an anomaly feature map. The final anomaly response values ​​satisfy the expression:

[0066] ;

[0067] in, For pixels The final abnormal response value; For pixels The initial abnormal response value; For pixels The model confidence at that point. When the model confidence... When the confidence level is higher, anomalous signals are preserved. The smaller the value, the more suppressed the abnormal signal becomes, ensuring that only abnormal signals appearing in reliable areas of the background model are identified as real defects, which can effectively reduce the false alarm rate.

[0068] S5. Locate the broken strand defect based on the abnormal feature map.

[0069] Specifically, the abnormal feature map is segmented using the Otsu threshold to obtain a binary defect image. Connectivity analysis is then performed on the binary defect image to obtain multiple connected components. For each connected component, the area of ​​the connected component is less than... Greater than This connected component is considered as the region of strand breakage defect. A preset minimum area threshold is used to filter out isolated noise pixels. In this embodiment, the threshold is... Set to 5 pixels in size; A preset maximum area threshold is used to filter out excessively large areas caused by other non-fractured strand anomalies. In this embodiment, [the threshold is set to be filled in]. The size is set to 500 pixels. In other embodiments, the implementer can set it according to the actual implementation situation. as well as .

[0070] For example, Figure 4 This is the result of the defect detection for broken strands. Figure 4 The area within the red box indicates a broken strand defect.

[0071] This invention also discloses an image-based overhead strand breakage detection system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an image-based overhead strand breakage detection method according to the present invention.

[0072] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

Claims

1. A method for detecting broken strands in overhead stranded wires based on images, characterized in that, include: Multi-scale and multi-directional decomposition is performed on the grayscale image of the overhead stranded line to obtain the original response components of the grayscale image in multiple scales and directions. Based on the local texture characteristics of grayscale images, a local harmonic model is constructed for each pixel in the grayscale image. This includes: constructing a structure tensor based on the gradient information of each pixel in the grayscale image; performing eigenvalue decomposition on the structure tensor to obtain the optimized local texture direction; determining the direction index with the smallest angular difference from the optimized local texture direction among multiple directions; obtaining the amplitude of the pixel at the direction index at multiple scales; obtaining the scale corresponding to the largest amplitude; and using the center frequency of the Log-Gabor filter corresponding to that scale as the local texture frequency of the pixel; combining the optimized local texture direction and the local texture frequency to form the local harmonic model of the pixel; the local harmonic model is used to characterize the direction and frequency characteristics of the normal line texture in the region where the pixel is located. Based on the local harmonic model and the multi-scale, multi-directional original response components, the non-harmonic weights of each original response component are calculated. The non-harmonic weights are used to characterize the degree of deviation between the original response components and the local harmonic model. An anomaly feature map is constructed by weighting the original response components of multiple scales and directions according to non-harmonic weights. This includes: weighting the original response components of multiple scales and directions according to non-harmonic weights to obtain the initial anomaly response value of the pixel; constructing the model confidence of the pixel based on the magnitude of the pixel at all scales along the direction index; multiplying the model confidence by the initial anomaly response value as the final anomaly response value of the pixel; and constructing the anomaly feature map from the final anomaly response values ​​of all pixels. Based on the abnormal feature map, locate the stock breakage defect; The process of obtaining the optimized local texture direction includes: decomposing the original local texture direction at a pixel into horizontal vector components and vertical vector components; performing Gaussian smoothing on the horizontal and vertical vector components of all pixels to obtain smoothed horizontal and vertical vector components; and combining the smoothed horizontal and vertical vector components to obtain the optimized local texture direction. The nonharmonic weights satisfy the expression: It should be noted that, For pixels In scale and direction Nonharmonic weights; For the index of the preset direction, For pixels Optimize local texture direction, For direction and optimize local texture direction The minimum angular difference between them; For scale The corresponding center frequency of the Log-Gabor filter; For pixels Local texture frequencies; It is the natural logarithm function; It is a natural exponential function; These are directional bandwidth control parameters; These are frequency bandwidth control parameters.

2. The image-based method for detecting broken strands in overhead stranded wires according to claim 1, characterized in that, Multi-scale, multi-directional decomposition is performed on the grayscale image of the overhead stranded line to obtain the multi-scale, multi-directional original response components of the grayscale image, including: A multi-scale, multi-directional Log-Gabor filter bank is used to filter the grayscale image of the overhead stranded line to obtain the complex response matrix at each scale and in each direction. The amplitude component and the phase component are separated from the complex response matrix at each scale and in each direction as the original response components at each scale and in each direction.

3. The image-based method for detecting broken strands in overhead stranded wires according to claim 1, characterized in that, The initial anomaly response value satisfies the expression: ； in, For pixels The initial abnormal response value; For pixels In scale and direction Nonharmonic weights; For pixels In scale and direction The range of the above; Represents pixels In scale and direction Phase on, It is a cosine function. It is a sine function; Indicates the scale quantity. Indicates the number of preset directions; This means that only positive values ​​are considered, and negative values ​​are counted as 0. To prevent extremely small positive numbers with a denominator of zero.

4. The image-based method for detecting broken strands in overhead stranded wires according to claim 1, characterized in that, The confidence level of the model satisfies the expression: ； in, For pixels Model confidence at the location; For pixels In scale and direction The range of the upper part, for In each preset direction and pixel point The index of the direction with the smallest local texture direction angle difference is optimized. For pixels In direction The index of the scale corresponding to the largest amplitude among all scales; For pixels In scale and direction The range of the upper part, For pixels In direction The sum of the amplitudes of all scales above, Indicates the scale quantity. Indicates the number of preset directions.

5. The image-based method for detecting broken strands in overhead stranded cables according to claim 1, characterized in that, Based on the aforementioned abnormal feature map, the defect of broken stock is located, including: The abnormal feature map is segmented by Otsu thresholding to obtain a defect binary image; the defect binary image is analyzed by connected component analysis to obtain multiple connected components; in response to the connected component area being greater than a preset minimum area threshold and less than a preset maximum area threshold, the connected component is taken as the broken strand defect region.

6. An image-based overhead strand breakage detection system, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement an image-based method for detecting broken strands in overhead stranded wires according to any one of claims 1-5.

Images