Power transmission line channel positive sample feature mining comparison method and device based on EMA

Abstract

The invention relates to the field of intelligent visual monitoring and industrial image feature mining, and provides a power transmission line channel positive sample feature mining comparison method and device based on EMA. According to the method, the accuracy and efficiency of power transmission ground channel anomaly detection are improved by establishing a positive sample feature fusion library, using a YOLO classification model and an exponential moving average algorithm and focusing on time sequence feature differences of point historical analysis data. According to the technical scheme, instance segmentation, feature extraction and classification are carried out on an inspection image, an exponential decay fusion mechanism is used for updating a feature fusion library, change of feature values is judged by setting a truncation threshold value, recognition of a stable change area is achieved, false detection is filtered through feature matching and Pearson's correlation coefficient analysis, and the model reporting precision is improved. Compared with the prior art, the method has the advantages that the detection precision and adaptability are improved, the manual recheck burden is reduced, the model learning is optimized, and the method has remarkable technical advantages and application value.

Description

Technical Field

[0001] The present invention relates to the fields of intelligent visual monitoring and industrial image feature mining, and mainly provides a method and device for mining and comparing positive sample features of an EMA power transmission line channel. Background Art

[0002] In the current technical field, the detection of foreign objects in the ground channel of a power transmission line is a key issue, which is directly related to the stable operation of the power grid and power supply safety. Existing technical solutions usually include collecting a large amount of abnormal sample data of the power transmission ground channel, classifying and labeling the known abnormal data samples, and then training a target detection model for detection.

[0003] However, there are some problems with existing technical solutions. First, due to the particularity of the power transmission channel scene, it is a very difficult process to obtain a large amount of data for model training, which leads to the incompleteness of sample data, thus affecting the accuracy of traditional negative sample detection models. Second, the ground scene of the power transmission channel has a large span, a high degree of complexity in open scenes, and many types of abnormal situations, making it difficult to achieve complete coverage of all abnormal types. Conventional detection models generally have relatively serious problems such as missed detection and false detection. Finally, due to the actual supervision requirements of the power transmission business, alarms are only issued when there are drastic changes in the foreign object form (shape, position, area, etc.). However, traditional detection algorithms alarm for all detected foreign objects and cannot classify the alarm situations, increasing the burden of manual review. Summary of the Invention

[0004] The purpose of the present invention is to solve the challenges and problems posed by numerous unforeseeable foreign object types due to environmental factors such as wide span and complex scenes in the abnormal detection of the power transmission ground channel, and the professional requirement of only reporting stable foreign object changes. By establishing a sample feature fusion library and using a method that combines the YOLO classification model with the exponential smoothing average algorithm, without concerning the specific abnormal categories and quantities, only focusing on the time series feature differences between each point and its historical analysis data, when the cumulative degree of the differences meets the specified threshold range, it is regarded as an abnormal change and an alarm is issued, thereby improving the accuracy and efficiency of abnormal detection.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] The present invention provides a method for mining and comparing positive sample features of an EMA power transmission line channel, including the following steps:

[0007] Step 1: Detect and identify the ground abnormal detection images of the power transmission channel, extract the features of the detected foreign objects, and establish a positive sample feature fusion library;

[0008] Step 2: Detect and identify the time-series images sampled at the same position, extract the foreign object features, perform weighted averaging according to the time sequence, and determine whether there is an abnormal feature change at the current moment, and alarm the foreign object instances with changes;

[0009] Step 3: Perform instance feature matching on the abnormal instances in Step 2, further filter out the false detections caused by light and shadow and background changes, improve the reporting accuracy of the model, and at the same time distill and purify the positive sample features in the feature fusion library to facilitate the subsequent learning and judgment of the model.

[0010] Further, Step 1 includes the following steps:

[0011] Step 1.1: Perform instance segmentation on the inspection images, extract the feature information in each instance in a way that is accurate to the target pixels to obtain instance features. An instance refers to a specific area or object identified as a foreign object in the image;

[0012] Step 1.2: Centering on the instance obtained by segmentation, extend about 5% of the instance pixel area around it, and inject the extended background information into the instance features to assist the model in judgment;

[0013] Step 1.3: Use the trained classification model to classify the instance features obtained in Step 1.2. The instances that need to be detected are classified into classes with fixed names, called foreign object instances, filter out the known normal situation instances, and the remaining instances are classified into the suspected abnormal class, called suspected foreign object instances;

[0014] Step 1.4: Perform exponential decay fusion on the feature information of all instances, and then store it in the feature fusion library. Initialize the feature fusion library. The exponential decay fusion of features needs to satisfy the following formula:

[0015]

[0016] where: y represents the feature value extracted from the current input image, xt +1 represents the feature value after experiencing exponential decay fusion, x t represents the feature value cached in the current feature library, α represents the decay coefficient with a value between [0, 1], and the decay fusion ratio of historical features and current features is controlled by adjusting the value of α to adapt to different point inspection frequencies.

[0017] Further, Step 2 includes the following steps:

[0018] Step 2.1: Analyze the pictures in chronological order, use the formula in Step 1.4 for feature fusion, and update the feature fusion library;

[0019] Step 2.2: Iterate step 2.1, continuously update the feature fusion library, and determine whether the feature values in the feature library meet two truncation thresholds, α and β, where α represents the new truncation threshold, marking the gradual change of foreign object instances in the time series; β represents the disappearance truncation threshold, marking the gradual disappearance of foreign object instances in the time series;

[0020] Step 2.3: Divide the feature fusion library into three feature regions according to the truncation thresholds:

[0021] The region where the feature value is between [0, α] is the unstable change region. The feature changes in this region are probably caused by the image light and shadow background or the unstable model detection results;

[0022] The region where the feature value is between [α, β] is the stable change region. The feature changes in this region are probably caused by the real abnormal changes of the detection target and are the main reference indicators for alarms;

[0023] The region where the feature value is between [β, max_value] is the stable region. There are no abnormal feature changes in this region, which is the region where foreign object instances stably exist in the current image;

[0024] Step 2.4: Quantize and save the feature values obtained in step 2.3 to compress the storage space and speed up data reading and writing.

[0025] Furthermore, perform feature distillation and purification through step 3. The specific steps include the following:

[0026] Step 3.1: Select a time neighborhood according to the α and β thresholds. The picture sequence within the time neighborhood is regarded as forming a time window, and the first picture within the time window is the reference picture with the largest accumulation of abnormal change features;

[0027] Step 3.2: Extract the abnormal feature stable change region in step 2.3 and obtain the original pixel values of this region on the current image;

[0028] Step 3.3: Normalize the pixel values in the pixel neighborhood corresponding to the current detection picture and the reference picture, and use the Pearson correlation coefficient to judge the similarity between the two. Selecting a suitable pixel neighborhood (generally 0.05 times the side length of the target, or fixed at a dozen or dozens of pixels according to the actual situation) can alleviate the shooting error caused by camera jitter.

[0029] Step 3.4: If the similarity in step 3.3 is greater than the set threshold, take out the outermost polygon of each contour point set of the foreign object instance contour point set in the detection area as the reported abnormal output result, and update the abnormal change region in the feature fusion library with a relatively large weight.

[0030] Step 3.5: If the similarity degree in Step 3.3 does not reach the set threshold, reset the feature values of this area in the feature fusion library according to the change trend of the characteristics of the abnormal change area.

[0031] The present invention also provides a device for mining and comparing positive sample features of an EMA transmission line channel, including:

[0032] A detection and recognition module, which is used to detect and recognize the ground abnormal detection images of the transmission channel, extract the features of the detected foreign objects, and establish a positive sample feature fusion library;

[0033] A time series analysis module, which is used to detect and recognize the time series pictures obtained by sampling at the same position, extract the foreign object features, perform weighted averaging according to the time series, and judge whether there is an abnormal feature change at the current moment, and alarm the foreign object instances with changes;

[0034] An instance matching module, which is used to perform instance feature matching on the abnormal instance reports in the time series analysis module, further filter out the false detections caused by light and shadow and background changes, improve the reporting accuracy of the model, and at the same time distill and purify the positive sample features in the feature fusion library, which is beneficial to the subsequent learning and judgment of the model.

[0035] In the above device, the detection and recognition module includes:

[0036] An instance segmentation unit, which is used to perform instance segmentation on the inspection images, extract the feature information in each instance in a way that is accurate to the target pixels, and obtain instance features;

[0037] A background information injection unit, which is used to extend about 5% of the instance pixel area around the instance obtained by segmentation, and inject the extended background information into the instance features to assist the model in judgment;

[0038] A classification model unit, which is used to classify the instance features obtained in the background information injection unit using the trained classification model. The instances that need to be detected are classified into classes with fixed names, called foreign object instances, filter out the known normal situation instances, and the remaining instances are classified into the suspected abnormal class, called suspected foreign object instances;

[0039] A feature fusion library initialization unit, which is used to perform exponential decay fusion on the feature information of all instances, then store it in the feature fusion library, and perform an initialization operation on the feature fusion library.

[0040] In the above device, the time series analysis module includes:

[0041] A feature fusion update unit, which is used to analyze the pictures in chronological order, perform feature fusion using the formula in the feature fusion library initialization unit, and update the feature fusion library;

[0042] A feature threshold judgment unit for iterating the feature fusion update unit, continuously updating the feature fusion library, and judging whether the feature values in the feature library meet two truncation thresholds, α and β;

[0043] A feature region division unit for dividing the feature fusion library into three feature regions according to the truncation threshold;

[0044] A feature quantization and storage unit for quantizing and storing the feature values obtained in the feature region division unit, compressing the storage space, and accelerating the data reading and writing speed.

[0045] In the above device, the instance matching module includes:

[0046] A time window selection unit for selecting a time neighborhood according to the α and β thresholds, and regarding the picture sequence within the time neighborhood as forming a time window;

[0047] An abnormal feature extraction unit for extracting the stable change region of abnormal features in the feature quantization and storage unit, and obtaining the original pixel values of this region on the current image;

[0048] A pixel similarity judgment unit for normalizing the pixel values in the pixel neighborhood corresponding to the reference image in the time window selection unit with the current detection image, and using the Pearson correlation coefficient to judge the similarity between the two;

[0049] An abnormal output update unit for, if the similarity degree in the pixel similarity judgment unit is greater than the set threshold, taking the outermost polygon of each contour point set of the foreign object instance contour point set within the detection region as the reported abnormal output result, and giving a relatively large weight to update the abnormal change region in the feature fusion library;

[0050] A feature value reset unit for, if the similarity degree in the pixel similarity judgment unit does not reach the set threshold, resetting the feature values of this region in the feature fusion library according to the feature change trend of the abnormal change region.

[0051] The present invention provides a method for mining and comparing positive sample features of a transmission line channel based on EMA (Exponential Moving Average), aiming to improve the accuracy and efficiency of abnormal detection of the transmission line ground channel. This method realizes precise warning of stable foreign object changes by establishing a sample feature fusion library and using the YOLO classification model and the exponential moving average algorithm, focusing on the time series feature differences between each point and its historical analysis data.

[0052] Analysis of technical effects:

[0053] Establishment of the positive sample feature fusion library:

[0054] By performing instance segmentation on the inspection images, accurately extracting the foreign object features, and establishing a positive sample feature fusion library.

[0055] The exponential decay fusion mechanism of the features ensures the reasonable fusion of historical features and current features, adapting to the inspection frequencies at different positions.

[0056] Time series feature analysis and anomaly detection:

[0057] Using the exponential moving average algorithm to perform feature fusion on the time series images and update the feature fusion library.

[0058] By setting two truncation thresholds, α and β, to judge the changes in the feature values in the feature library, realizing the identification of the stable change area and improving the accuracy of the alarm.

[0059] Feature distillation and false alarm filtering:

[0060] Through feature matching and Pearson correlation coefficient analysis, filtering out false detections caused by light and shadow and background changes.

[0061] Performing distillation on the feature fusion library to enhance the model's learning and judgment ability for abnormal features.

[0062] Advantages compared with the existing technology:

[0063] Improving the detection accuracy: Through feature fusion and the exponential moving average algorithm, reducing missed detections and false detections, and improving the accuracy of the detection model.

[0064] Strong adaptability: The exponential decay fusion mechanism of the features can adapt to different inspection frequencies and improve the adaptability of the model.

[0065] Reducing the manual review burden: Accurate foreign object change alarms reduce the need for manual review of all detected foreign objects, improving work efficiency.

[0066] Optimizing model learning: The feature distillation step enhances the model's learning ability for abnormal features and continuously improves the detection accuracy.

[0067] In summary, through innovative feature mining and comparison methods, the present invention effectively improves the accuracy and efficiency of abnormal detection in the transmission line corridor, while reducing the workload of manual review, having significant technical advantages and application values. Description of the Drawings

[0068] Figure 1 It is the overall flowchart of the algorithm framework;

[0069] Figure 2 It is the update flowchart of the positive sample feature fusion library. Detailed Implementation Modes

[0070] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0071] In addition, for a better illustration of the present invention, numerous specific details are given in the following specific embodiments. Those skilled in the art will understand that the present invention can be implemented without these specific details.

[0072] Embodiment 1

[0073] The present invention provides a method for mining and comparing positive sample features of EMA transmission line channels, including the following steps:

[0074] Step 1: Detect and identify the ground anomaly detection images of the transmission channel, extract the features of the detected foreign objects, and establish a positive sample feature fusion library.

[0075] Step 2: Detect and identify the time-series pictures sampled at the same location, extract the foreign object features, perform weighted averaging according to the time series, and determine whether there is an abnormal feature change at the current moment, and alarm the foreign object instances with changes.

[0076] Step 3: Perform instance feature matching on the abnormal reports in Step 2, further filter out the false detections caused by light and shadow and background changes, improve the reporting accuracy of the model, and at the same time distill and purify the positive sample features in the feature fusion library, which is beneficial for the subsequent learning and judgment of the model.

[0077] In the above technical solution, in order to solve the problem of extracting the features of all types of foreign objects for comparison, the main process of extracting foreign object features is established through Step 1, and a positive sample feature fusion library and a feature index decay fusion mechanism are established. At each subsequent detection moment, the foreign object feature changes will be compared based on the updated and fused feature library. The specific steps include the following content:

[0078] Step 1.1: Perform instance segmentation on the inspection images, and extract the feature information in each instance in a way that is accurate to the target pixels to obtain instance features. An instance refers to a specific area or object in the image that is recognized as a foreign object.

[0079] Step 1.2: Centering on the instance obtained by segmentation, extend about 5% of the instance pixel area around it, and inject the background information obtained by the extension into the instance features to assist the model in making judgments.

[0080] Step 1.3: Classify the instance features obtained in Step 1.2 using the trained classification model. The instances to be detected are classified into classes with fixed names, called foreign object instances, and the known normal instances are filtered out. The remaining instances are classified into the suspected abnormal class, called suspected foreign object instances.

[0081] Step 1.4: Perform exponential decay fusion on the feature information of all instances, and then store it in the feature fusion library and initialize the feature fusion library. The exponential decay fusion of features needs to satisfy the following formula:

[0082]

[0083] where: y represents the feature value extracted from the current input image, x t+1 represents the feature value after experiencing exponential decay fusion, x t represents the feature value cached in the current feature library, and α represents the decay coefficient with a value between [0, 1]. By adjusting this value, the decay fusion ratio of historical features and current features can be controlled to adapt to different site inspection frequencies.

[0084] In the above technical solution, in order to judge the range and trend of changes in foreign objects, the main process of comparing the changes in foreign object features is established through Step 2. By dividing different feature regions, the stable change trend of foreign objects is judged, and the area enclosed by the feature regions represents the change range of foreign objects. The specific steps include the following:

[0085] Step 2.1: Analyze the pictures in chronological order, perform feature fusion using the formula in Step 1.4, and update the feature fusion library;

[0086] Step 2.2: Iterate Step 2.1, continuously update the feature fusion library, and judge whether the feature values in the feature library meet the two truncation thresholds α and β. Among them, α represents the new truncation threshold, indicating the changes in foreign object instances gradually appearing in the time series; β represents the disappearance truncation threshold, indicating the changes in foreign object instances gradually disappearing in the time series.

[0087] Step 2.3: Divide the feature fusion library into three feature regions according to the truncation thresholds (classify the feature values within the same segment into one category, and a total of three categories are divided),

[0088] The region where the feature value is between [0, a] is the unstable change region. The feature changes in this region are probably caused by unstable image light and shadow background or unstable model detection results.

[0089] The region where the feature value is between [α, β] is the stable change region. The feature changes in this region are probably caused by real abnormal changes in the detection target and are the main reference indicators for alarm.

[0090] The region where the eigenvalue is between [β, max_value] is the stable region. There is no abnormal feature change in this region, which is the region where foreign object instances stably exist in the current image.

[0091] Step 2.4: Quantize and save the eigenvalues obtained in Step 2.3 to compress the storage space and accelerate the data reading and writing speed.

[0092] In the above technical solution, in order to further filter the reports generated by model misdetection, distill the eigenvalues in the feature library, and improve the detection accuracy and the learning ability of the model, feature purification is performed through Step 3. The specific steps include the following:

[0093] Step 3.1: Select a time neighborhood according to the α and β thresholds. The picture sequence within the time neighborhood is regarded as forming a time window, and the first picture within the time window is the reference picture with the largest cumulative abnormal change feature.

[0094] Step 3.2: Extract the stable change region of the abnormal feature in Step 2.3, and obtain the original pixel values of this region on the current image.

[0095] Step 3.3: Normalize the pixel values in the pixel neighborhood corresponding to the current detection image and the reference image, and use the Pearson correlation coefficient to judge the similarity between the two. Selecting an appropriate pixel neighborhood (generally 0.05 times the side length of the target) can alleviate the shooting error caused by camera jitter.

[0096] Step 3.4: If the similarity in Step 3.3 is greater than the set threshold, the contour point set of the foreign object instance in the detection area is taken out, and the outermost polygon of each contour point set is used as the reported abnormal output result, and a relatively large weight is given to update the abnormal change area in the feature fusion library.

[0097] Step 3.5: If the similarity in Step 3.3 does not reach the set threshold, the eigenvalues of this region in the feature fusion library are reset according to the feature change trend of the abnormal change region (the eigenvalue refers to a number between 0 and 255, which is used to represent the feature accumulation intensity and change trend at the current position. Resetting the eigenvalue is to reset it to 255 or 0 according to the change rule of the eigenvalue). This step mainly filters false alarms of warning information and redistills and purifies part of the content in the feature fusion library, so that the model can better learn the deep meaning of abnormal feature changes and improve the prediction accuracy.

[0098] Embodiment 2

[0099] The present invention also provides a device for mining and comparing positive sample features of an EMA transmission line channel, including:

[0100] The detection and recognition module is used to detect and recognize the ground anomaly detection images of the transmission channel, extract the features of the detected foreign objects, and establish a positive sample feature fusion library;

[0101] The time series analysis module is used to detect and recognize the time series images obtained by sampling at the same point, extract the features of foreign objects, perform weighted averaging according to the time series, and judge whether there is an abnormal feature change at the current moment, and alarm the foreign object instances with changes;

[0102] The instance matching module is used to perform instance feature matching on the abnormal instance reports in the time series analysis module, further filter out the false detections caused by light and shadow and background changes, improve the reporting accuracy of the model, and at the same time distill and purify the positive sample features in the feature fusion library, which is beneficial to the subsequent learning and judgment of the model.

[0103] In the above device, the detection and recognition module includes:

[0104] The instance segmentation unit is used to perform instance segmentation on the inspection images, extract the feature information in each instance in a way that is accurate to the target pixels, and obtain the instance features;

[0105] The background information injection unit is used to extend about 5% of the instance pixel area around the segmented instance, and inject the extended background information into the instance features to assist the model in judgment;

[0106] The classification model unit is used to classify the instance features obtained in the background information injection unit using the trained classification model. The instances that need to be detected are classified into classes with fixed names, called foreign object instances, filter out the known normal situation instances, and classify the remaining instances into the suspected abnormal class, called suspected foreign object instances;

[0107] The feature fusion library initialization unit is used to perform exponential decay fusion on the feature information of all instances, and then store it in the feature fusion library to perform initialization operations on the feature fusion library.

[0108] In the above device, the time series analysis module includes:

[0109] The feature fusion update unit is used to analyze the pictures in chronological order, perform feature fusion using the formula in the feature fusion library initialization unit, and update the feature fusion library;

[0110] The feature threshold judgment unit is used to iterate on the feature fusion update unit, continuously update the feature fusion library, and judge whether the feature values in the feature library meet the two cut-off thresholds of α and β;

[0111] The feature region division unit is used to divide the feature fusion library into three feature regions according to the cut-off threshold;

[0112] A feature quantization and storage unit, which is used to quantize and store the feature values obtained by the feature region division unit, compress the storage space, and accelerate the data reading and writing speed.

[0113] In the above-mentioned device, the instance matching module includes:

[0114] A time window selection unit, which is used to select a time neighborhood according to the α and β thresholds, and the picture sequence within the time neighborhood is regarded as forming a time window;

[0115] An abnormal feature extraction unit, which is used to extract the stable change region of the abnormal feature in the feature quantization and storage unit, and obtain the original pixel values of this region on the current image;

[0116] A pixel similarity judgment unit, which is used to normalize the pixel values in the pixel neighborhood corresponding to the reference image in the time window selection unit and the current detection image, and use the Pearson correlation coefficient to judge the similarity between the two;

[0117] An abnormal output update unit, which is used to, if the similarity degree in the pixel similarity judgment unit is greater than the set threshold, take the outermost polygon of each contour point set of the foreign object instance contour point set within the detection area as the reported abnormal output result, and give a relatively large weight to update the abnormal change region in the feature fusion library;

[0118] A feature value reset unit, which is used to, if the similarity degree in the pixel similarity judgment unit does not reach the set threshold, reset the feature values of this region in the feature fusion library according to the feature change trend of the abnormal change region.

[0119] Embodiment 3

[0120] The present invention also discloses a transmission ground channel positive sample detection device based on exponential moving average, including:

[0121] A graph feature extraction unit, which is used to extract features from the input image and classify the foreign object features that need to be concerned about.

[0122] A positive sample feature fusion unit, which fuses the time series features through an improved moving exponential average algorithm, and marks the abnormal change region according to the feature value range.

[0123] A reference image selection unit, which automatically calculates and selects the picture with the largest cumulative change of foreign object features in the time window neighborhood as the reference image through the abnormal change threshold.

[0124] An abnormal feature matching unit, which further filters false alarms by statistically analyzing the pixel distribution relationship between the abnormal change region in the feature library and the appropriate pixel neighborhood of the reference image and calculating the correlation coefficient between the feature values. Distill and purify the feature library to improve the positive sample feature learning ability of the model and the overall reporting accuracy of the algorithm.

[0125] In summary, this technology establishes a positive sample feature fusion library through an improved EMA technology and combines a semantic segmentation and classification model.

Claims

1. A method for mining and comparing positive sample features of EMA transmission line channels, characterized in that, It includes the following steps: Step 1: Detect and identify the ground anomaly detection images of the power transmission channel, extract the features of the detected foreign objects, and establish a positive sample feature fusion library; Step 2: Detect and identify the time series images sampled at the same location, extract the foreign object features, perform weighted averaging according to the time sequence, and judge whether there is an abnormal feature change at the current moment, and alarm the foreign object instances with changes; Step 3: Perform instance feature matching on the abnormal instances in Step 2 to further filter out the false detections caused by light and shadow and background changes, improve the reporting accuracy of the model, and at the same time distill and purify the positive sample features in the feature fusion library to facilitate the subsequent learning and judgment of the model.

2. The method for mining and comparing positive sample features of an EMA transmission line channel according to claim 1, wherein, Step 1 includes the following steps: Step 1.1: Perform instance segmentation on the inspection images, extract the feature information in each instance in a way that is accurate to the target pixels to obtain instance features. An instance refers to a specific area or object identified as a foreign object in the image; Step 1.2: Centering on the instance obtained by segmentation, extend about 5% of the instance pixel area around it, and inject the extended background information into the instance features to assist the model in judgment; Step 1.3: Use the trained classification model to classify the instance features obtained in Step 1.

2. The instances that need to be detected are classified into classes with fixed names, called foreign object instances, filter out the instances of known normal situations, and the remaining instances are classified into the suspected abnormal class, called suspected foreign object instances; Step 1.4: Perform exponential decay fusion on the feature information of all instances, and then store it in the feature fusion library. Initialize the feature fusion library. The exponential decay fusion of features needs to satisfy the following formula: Among them: y represents the feature value extracted from the current input image, x t+1 represents the eigenvalue after exponential decay fusion, x t It represents the feature value cached in the current feature library, and α represents the attenuation coefficient between [0, 1]. By adjusting the value of α, the attenuation fusion ratio of historical features and current features can be controlled to adapt to different point inspection frequencies.

3. A method for mining and comparing positive sample features of an EMA power transmission line channel according to claim 2, characterized in that, Step 2 includes the following steps: Step 2.1: Analyze the pictures in chronological order, perform feature fusion using the formula in Step 1.4, and update the feature fusion library; Step 2.2: Iterate Step 2.1, continuously update the feature fusion library, and judge whether the feature values in the feature library meet two truncation thresholds, α and β. Among them, α represents the new addition truncation threshold, indicating the changes of foreign object instances gradually appearing in the time sequence; β represents the disappearance truncation threshold, indicating the changes of foreign object instances gradually disappearing in the time sequence; Step 2.3: Divide the feature fusion library into three feature regions according to the truncation threshold: The region where the feature value is between [0, a] is the unstable change region. The feature changes in this region are probably caused by the instability of the image light and shadow background or the model detection result; The region where the feature value is between [α, β] is the stable change region. The feature changes in this region are probably caused by the real abnormal changes of the detection target, and it is the main reference index for alarming; The region where the feature value is between [β, max_value] is the stable region. There are no abnormal feature changes in this region, and it is the region where foreign object instances stably exist in the current image; Step 2.4: Quantize and save the feature values obtained in Step 2.3 to compress the storage space and speed up the data reading and writing speed.

4. A method for mining and comparing positive sample features of an EMA power transmission line channel according to claim 3, characterized in that, Feature distillation and purification are carried out through Step 3. The specific steps include the following contents: Step 3.1: Select a temporal neighborhood according to the α and β thresholds. The image sequence within the temporal neighborhood is regarded as constituting a temporal window, and the first image within the temporal window is the reference image with the maximum cumulative abnormal change features. Step 3.2: Extract the stable change region of the abnormal features in Step 2.3, and obtain the original pixel values of this region on the current image. Step 3.3: Normalize the pixel values in the pixel neighborhood of the corresponding regions of the current detection image and the reference image, and use the Pearson correlation coefficient to judge the similarity between the two. Step 3.4: If the similarity in Step 3.3 is greater than the set threshold, the contour point set of the foreign object instance within the detection region is taken, and the outermost polygon of each contour point set is used as the reported abnormal output result, and a relatively large weight is given to update the abnormal change region in the feature fusion library. Step 3.5: If the similarity in Step 3.3 does not reach the set threshold, reset the feature values of this region in the feature fusion library according to the change trend of the abnormal change region features.

5. A device for mining and comparing positive sample features of an EMA power transmission line channel, characterized in that Including: A detection and recognition module, which is used to detect and recognize the abnormal detection images of the transmission channel ground, extract the detected foreign object features, and establish a positive sample feature fusion library. A time series analysis module, which is used to detect and recognize the time series images obtained by sampling at the same position, extract foreign object features, perform weighted averaging according to the time series, and judge whether abnormal feature changes occur at the current moment, and alarm the foreign object instances with changes. An instance matching module, which is used to perform instance feature matching on the abnormal instance reports in the time series analysis module, further filter out false detections caused by light and shadow and background changes, improve the model reporting accuracy, and at the same time distill and purify the positive sample features in the feature fusion library to facilitate the subsequent learning and judgment of the model.

6. The device according to claim 5, characterized in that, The detection and recognition module includes: An instance segmentation unit, which is used to perform instance segmentation on the inspection images, extract the feature information in each instance in a way that is accurate to the target pixels, and obtain instance features. A background information injection unit, which is used to extend about 5% of the instance pixel area around the instance obtained by segmentation, and inject the extended background information into the instance features to assist the model in judgment. A classification model unit, which is used to classify the instance features obtained in the background information injection unit using the trained classification model. The instances that need to be detected are classified into classes with fixed names, called foreign object instances, filter out the known normal situation instances, and the remaining instances are classified into the suspected abnormal class, called suspected foreign object instances. A feature fusion library initialization unit, which is used to perform exponential decay fusion on the feature information of all instances, and then store it in the feature fusion library, and perform initialization operations on the feature fusion library.

7. The device according to claim 6, characterized in that The time series analysis module includes: A feature fusion update unit, which is used to analyze the pictures in chronological order, perform feature fusion using the formula in the feature fusion library initialization unit, and update the feature fusion library. A feature threshold judgment unit, which is used to iterate on the feature fusion update unit, continuously update the feature fusion library, and judge whether the feature values in the feature library meet the two truncation thresholds α and β. A feature region division unit, which is used to divide the feature fusion library into three feature regions according to a truncation threshold; A feature quantization and storage unit, which is used to quantize and store the feature values obtained in the feature region division unit, compress the storage space, and accelerate the data reading and writing speed.

8. The device according to claim 7, characterized in that The instance matching module includes: A time window selection unit, which is used to select a time neighborhood according to the α and β thresholds, and the image sequence within the time neighborhood is regarded as constituting a time window; An abnormal feature extraction unit, which is used to extract the stable change region of the abnormal feature in the feature quantization and storage unit, and obtain the original pixel values of this region on the current image; A pixel similarity judgment unit, which is used to normalize the pixel values in the pixel neighborhood corresponding to the reference image in the time window selection unit for the current detection image, and use the Pearson correlation coefficient to judge the similarity between the two; An abnormal output update unit, which is used to, if the similarity in the pixel similarity judgment unit is greater than the set threshold, take the outermost polygon of each contour point set from the contour point sets of the foreign object instances in the detection area as the reported abnormal output result, and give a relatively large weight to update the abnormal change region in the feature fusion library; A feature value reset unit, which is used to, if the similarity in the pixel similarity judgment unit does not reach the set threshold, reset the feature values of this region in the feature fusion library according to the feature change trend of the abnormal change region.

Images