A method and system for optimizing false alarm rate of multi-model cascade transmission line defect identification

Abstract

The present invention discloses a multi-model cascade transmission line defect recognition false alarm rate optimization method and system, the method calls a defect recognition model to perform defect recognition on a patrol image, calls a corresponding negative sample recognition model to perform negative sample recognition according to the defect type, and obtains a negative sample recognition model research and judgment result; compares the negative sample recognition model research and judgment result with the defect type and defect position extracted by the defect recognition model to determine whether it is a false alarm defect; if it is not a false alarm defect, the defect is put into the output defect information, and if it is a false alarm defect, the defect is removed; traverses all defects in the patrol image and calls different negative sample recognition models respectively, conducts comprehensive research and judgment according to the output defect information, and finally forms a comprehensive research and judgment result. The present invention realizes the pressure drop of the overall false alarm rate by supplementing the negative sample recognition model and the multi-model cascade system integration means, which can effectively reduce the false alarm rate of image recognition of the unmanned aerial vehicle system.

Description

Technical Field

[0001] The present invention belongs to the technical field of power transmission line defect detection, and relates to a method and system for optimizing the false alarm rate of multi-model cascade power transmission line defect identification. Background Art

[0002] As a new mode of digital transformation and operation and maintenance of power transmission that replaces manual inspection, autonomous inspection by drones has attracted much attention since its birth. Autonomous inspection by drones can realize the digital online management of the entire business flow of plan preparation-task formulation-task dispatch-inspection operation-safety monitoring-data feedback-image recognition-defect flow, and the image recognition technology based on artificial intelligence is the core technical means in this business process. Through this technical means, the time of manual defect identification can be greatly reduced.

[0003] The defect recognition model of the current artificial intelligence platform is mainly constructed by training defect samples (positive samples). However, the image defect scenes are highly complex, and the defect samples cannot be fully covered. It is easy to misjudge normal targets as abnormal targets. For example, the spacer rod on the conductor is identified as a foreign object on the conductor. As a result, the grassroots team needs to spend a lot of manpower to verify and edit the misjudged defects, which greatly affects the level of automation in the autonomous inspection process. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method and system for optimizing the false alarm rate of defect identification of multi-model cascade transmission lines.

[0005] The present invention is implemented by the following technical solutions: A multi-model cascade transmission line defect identification false alarm rate optimization method, comprising:

[0006] Step S1: receiving patrol pictures;

[0007] Step S2: calling the defect recognition model to perform defect recognition on the inspection image, and extracting the defect type and defect location of each defect;

[0008] Step S3: Call the corresponding negative sample recognition model according to the defect type to perform negative sample recognition and obtain the negative sample recognition model judgment result;

[0009] Step S4: Compare the judgment result of the negative sample recognition model with the defect type and defect location extracted by the defect recognition model to determine whether it is a false alarm defect; if it is not a false alarm defect, put the defect into the output defect information; if it is a false alarm defect, remove the defect;

[0010] Step S5: Traverse all defects in the inspection image and call different negative sample recognition models respectively, conduct comprehensive analysis and judgment based on the output defect information, and finally form a comprehensive analysis and judgment result.

[0011] Further preferably, the specific process of step S3 is as follows:

[0012] Step S3-1: Enhance the patrol image according to the patrol features, and then use the enhanced image samples to train the negative sample recognition model;

[0013] Step S3-2: Realize the recognition of negative samples through the multi-overlapping negative sample recognition model calling strategy; crop the patrol image into several overlapping sub-images, and send the sub-images to the negative sample recognition model for recognition, merge the recognition target results of each sub-image in the negative sample recognition model, and then output the negative sample recognition model judgment result.

[0014] Further preferably, the enhancement processing includes contrast enhancement and backlight processing.

[0015] Further preferably, the contrast enhancement processing process is: firstly calculating the distribution of the image histogram, and then performing image histogram equalization processing based on the distribution of the image histogram.

[0016] Further preferably, the image histogram equalization processing steps are as follows:

[0017] According to the image histogram distribution, calculate the average brightness value;

[0018] Iterate over each pixel and increase the contrast by increasing the brightness difference between pixels.

[0019] Further preferably, the backlight processing includes: constructing illumination distribution templates of multiple directions and intensities, and applying the illumination distribution templates to the patrol picture.

[0020] Further preferably, for the patrol picture, the illumination distribution template and the patrol picture are fused by linear interpolation, and the illumination effect is simulated by superimposing the illumination distribution template in different areas.

[0021] Further preferably, the Yolox model is used as the negative sample recognition model.

[0022] Further preferably, LabelImg is used to annotate the negative sample defects.

[0023] Further preferably, in step S3-2, the recognition target results of each sub-image in the negative sample recognition model are merged:

[0024] Step S3-2-1: Traverse each sub-image after segmentation , q is the total number of sub-images after segmentation; for each sub-image Both call the negative sample recognition model for recognition and obtain the recognition target result set ,in For sub-image The nth recognition target information, e is the number of recognition targets, is the target type of the identified target, Is the recognition target in the sub-image The bounding box information on The definition is ,in The bounding boxes are in the sub-images The minimum value in the X-axis direction, the maximum value in the X-axis direction, the minimum value in the Y-axis direction, and the maximum value in the Y-axis direction;

[0025] Step S3-2-2: Collect each recognition target result The bounding box coordinates of the identified target are converted to the bounding box coordinates relative to the patrol image. The specific steps are: traverse the identified target result set Each identification target information in , read its bounding box information ,Will Add sub-picture The coordinates of the upper left corner on the patrol image ;

[0026] Step S3-2-3: Construct a set of final target recognition results ,in To finally identify the target information, is the target type of the final identification prediction target, is the bounding box information of the final recognition target recorded on the patrol image, and k is the number of recognition records;

[0027] Step S3-2-4: Traverse the sub-pictures one by one The recognition target result set ;for Each identification target information With each final recognition target information in the set N Make comparisons and determine the target type and Are they the same?

[0028] Step S3-2-5: If they are equal, merge and The bounding box of Get the merged bounding box The merging rules are in It is a record The bounding box parameters in , It is a record The bounding box parameters in ; and the bounding box Update to record of superior;

[0029] Step S3-2-6: If they are not equal, it means that it is a newly discovered prediction target, then record Put it into the set N of final target recognition results, and the set N of final target recognition results is the judgment result of the negative sample recognition model.

[0030] Further optimization is performed by calling different negative sample recognition models to conduct comprehensive research and judgment, and finally forming a comprehensive research and judgment result, including the following steps:

[0031] Step S5-1: First, after receiving the inspection picture and the corresponding defect set M extracted by the defect recognition model, the scheduling module searches one by one according to the defect type in the defect set M to see if there is a corresponding negative sample recognition model. If there is, these defects are extracted separately to form a set A, and the negative sample recognition model is called to perform a secondary judgment on the set A; the part of the defect set M excluding the set A is the set P;

[0032] Step S5-2: The scheduling module calls the negative sample recognition model to judge the set A, and forms a negative sample judgment result set B;

[0033] Step S5-3: After receiving the sets A, P, and B, the comprehensive analysis module conducts comprehensive analysis on the defects. The process is as follows:

[0034] Step S5-3-1: traverse each defect record a in set A to obtain a bounding box c;

[0035] Step S5-3-2: traverse the negative sample judgment result set B for the bounding box c, and determine whether the defect record in the negative sample judgment result set B intersects with c;

[0036] Step S5-3-3: If they intersect, remove the defect record a from set A;

[0037] Step S5-3-4: Determine whether all elements in set A have been traversed. If not, repeat steps S5-3-1 to S5-3-3 until all elements in set A have been traversed.

[0038] Step S5-3-5: After the traversal is completed, set A and set P are merged to form the final comprehensive analysis result set E.

[0039] The present invention also provides a multi-model cascade transmission line defect recognition false alarm rate optimization system, which includes three parts: a defect recognition model, a negative sample recognition model library, and a multi-model cascade system:

[0040] Defect recognition model: Preliminary identification of defect information in the inspection image, including defect type and defect location;

[0041] Multi-model cascade system: receives inspection images and defect information output by the defect recognition model; calls the corresponding negative sample recognition model from the negative sample recognition model library according to the defect type to conduct comprehensive analysis and judgment, eliminates erroneous defect information, and finally forms defect information after comprehensive analysis and judgment;

[0042] Negative sample recognition model library: includes several different negative sample recognition models, which are used to filter out defects that are misjudged by the defect recognition model.

[0043] The present invention can achieve a reduction in the overall false alarm rate by supplementing the negative sample recognition model and integrating the multi-model cascade system without changing the existing model, and can effectively reduce the false alarm rate of image recognition in the UAV system.

[0044] The present invention can greatly improve the efficiency of manual review during autonomous inspections, increase the degree of automation in defect identification during autonomous inspections, reduce the manpower cost investment required for defect review, and improve the practical level of UAV inspection image intelligent recognition technology, assisting grassroots teams in carrying out high-quality and high-precision autonomous inspection photography and defect identification and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The figure is a flow chart of the method of the present invention.

[0046] Figure 2 This is a schematic diagram for comprehensive analysis.

[0047] Figure 3 It is a system architecture diagram of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be further described in detail below with reference to the accompanying drawings.

[0049] like Figure 1 As shown, a multi-model cascade transmission line defect identification false alarm rate optimization method includes:

[0050] Step S1: receiving patrol pictures;

[0051] Step S2: Call the defect recognition model to perform defect recognition on the inspection image and extract the defect type and defect location of each defect. Usually the defect recognition model is provided externally;

[0052] Step S3: Call the corresponding negative sample recognition model according to the defect type to perform negative sample recognition and obtain the negative sample recognition model judgment result;

[0053] Step S4: Compare the judgment result of the negative sample recognition model (negative sample type and location) with the defect type and defect location extracted by the defect recognition model to determine whether it is a false alarm defect; if it is not a false alarm defect, put the defect into the output defect information; if it is a false alarm defect, remove the defect;

[0054] Step S5: Traverse all defects in the inspection image and call different negative sample recognition models respectively, conduct comprehensive analysis and judgment based on the output defect information, and finally form a comprehensive analysis and judgment result.

[0055] The specific process of step S3 in this embodiment is as follows:

[0056] Step S3-1: Enhance the inspection image according to the inspection features, and then use the enhanced image samples to train the negative sample recognition model; the inspection images are usually large in size, overexposed / underexposed, and shot against the light, which will interfere with the identification of inspection defects. Enhancement can improve the generalization ability of the negative sample recognition model. The sample enhancement includes contrast enhancement and backlight processing. Through these enhancement measures, the overexposure / underexposure, backlight and other situations in the inspection scene can be expanded;

[0057] Step S3-2: Since the negative sample recognition model can only receive images with a maximum resolution of 640x640, and the 4K patrol images collected by the drone must be downsampled, the resolution will be reduced from 4096x4096 to 640x640. In this downsampling process, small targets in the original image will be lost. Based on the large size of patrol images, this embodiment uses a multi-overlap negative sample recognition model calling strategy to achieve negative sample recognition without losing image resolution. The idea of ​​the multi-overlap negative sample recognition model calling strategy is to crop the patrol image into several sub-images with 50% overlap, and send the sub-images to the negative sample recognition model for recognition, and finally merge the recognition results for output.

[0058] For defects with high false alarm rates in the defect recognition model, the corresponding negative sample recognition model is analyzed and constructed (for example, foreign body defects in wires can easily misreport spacers and anti-vibration hammers as defects). The sample is expanded by adjusting the contrast parameters of the photos in the sample and simulating backlit shooting through an enhanced processing method based on the lighting characteristics in the patrol environment.

[0059] There are two situations in drone patrol images: overexposure and underexposure. When the photo is overexposed or underexposed, the overall brightness will be too high or too low. At this time, most pixels are concentrated in the brighter or darker areas, and the brightness difference between pixels is not large, resulting in unclear image details. This phenomenon can be improved through contrast enhancement processing measures.

[0060] Contrast enhancement mainly adjusts the image brightness globally based on the image brightness distribution statistics (i.e. histogram). The processing is divided into two steps:

[0061] 1. Calculate the image histogram distribution

[0062] Traverse each pixel, take the RGB value (R represents the red channel value, G represents the green channel value, and B represents the blue channel value), treat the RGB value as a three-dimensional space vector, calculate its brightness value, and finally get the brightness value of each pixel on the image. Sort the brightness values ​​from low to high, and count the number of times each brightness value appears, and finally get the distribution of each brightness value. The distribution set of brightness values ​​is the image histogram distribution.

[0063] 2. Based on the distribution of the image histogram, perform image histogram equalization processing

[0064] Because drone patrol pictures have two situations: overexposure and underexposure. These two situations cause the overall distribution of the image histogram to be concentrated in the dark or bright areas respectively. Therefore, it is necessary to dynamically consider the image histogram equalization processing measures according to the average brightness value of the image. When the image is underexposed, the average brightness value is low, the overall image is dark, and the brightness difference between adjacent pixels is not large, resulting in low contrast. It is necessary to brighten the image as a whole while increasing the brightness difference between adjacent pixels. On the contrary, when the image is overexposed, it is necessary to reduce the brightness of the overall image and increase the brightness difference between adjacent pixels. In summary, the steps of image histogram equalization are as follows:

[0065] 1) Calculate the average brightness value based on the image histogram distribution ,in is the brightness value contained in the image (0~255), is the brightness value The number of times it appears in the image.

[0066] 2) Traverse each pixel and increase the brightness difference between pixels to improve the contrast. The specific measures are to adjust the RGB value of each pixel according to the following rules:

[0067] when When , it means that the image is dark, and most of the pixels in the image are concentrated in the low-brightness area. The brightness difference between these pixels is small, resulting in unclear differences between pixels. Therefore, it is necessary to increase the brightness difference of these low-brightness pixels to achieve the effect of improving contrast. Formula The adjustment requirement can be met and the adjustment coefficient S is calculated; wherein r is the normalized brightness value and v is the overall magnification coefficient.

[0068] Assume that the brightness of two adjacent pixels is L1 and L2 respectively, and the brightness difference is , the brightness difference after curve conversion is , where K is The slope of the curve at K=1 is a critical point for brightness adjustment. When K is greater than 1, the brightness difference is greater than before adjustment, and when K is less than 1, the brightness difference is smaller than before adjustment. The change of the slope K of the curve is affected by v. The larger v is, the more obvious the adjustment is for the darker areas of the image. By adjusting the parameter v, the contrast between pixels with brightness lower than a specific value can be adjusted. According to the experience of this embodiment, the value of v is 255- .

[0069] when When , it means the image is bright, using the formula , calculate the adjustment coefficient S. When v is larger, the contrast adjustment of the brighter area of ​​the image is more obvious.

[0070] The new pixel RGB parameter value is adjusted to , . Represents the adjusted red channel value, Represents the adjusted green channel value, Represents the adjusted blue channel value.

[0071] When the drone takes photos against the light, in addition to the effect of overexposure, the sunlight itself will also cause great interference to the patrol pictures, thus causing misjudgment and missed judgment. In order to avoid misjudgment of the model, it is necessary to add the situation of taking photos against the light in the negative sample training. The backlight processing measures mainly include: constructing illumination distribution templates with multiple directions and intensities, and applying the illumination distribution templates to the patrol pictures.

[0072] Since color scattering and Tyndall effect will occur in backlit environments, it is impossible to simply simulate the lighting conditions through formula calculations. Therefore, the distribution diagram of sunlight in backlit photos at various angles is collected. By analyzing the image brightness, the effect of light brightness on backlit photos is extracted to form a lighting distribution template.

[0073] The specific extraction process of the illumination distribution template is as follows:

[0074] 1. Statistics of image brightness histogram.

[0075] 2. Separate the background and sunlight in backlit photos. Since the background accounts for a relatively large proportion of backlit photos, while the sunlight is relatively bright, the specific process is as follows:

[0076] Find the brightness mutation point. Since there is sunlight in the backlit photo, the brightness value of this part of the image is much higher than that of other images. Therefore, there will be a brightness mutation point in the highlight area (brightness value 200~255) in the histogram. Sort the brightness values ​​from high to low, calculate the difference in the number of pixels between two adjacent brightness values, and find the largest pixel difference. .in is the maximum pixel number difference, t is the brightness value, is the number of pixels with brightness value t, and T is the brightness when the maximum pixel number difference occurs.

[0077] Sort the brightness values ​​from low to high and determine the pixels in the range [0~T]. Because the brightness values ​​of these pixels are before the mutation point, they can be identified as background pixels. Strip out the brightness values ​​of background pixels and the brightness of illuminated pixels: first find all pixels with brightness lower than T, set the brightness of these pixels to 0, and obtain a background-stripped illuminated image, that is, the illumination distribution template.

[0078] For the patrol pictures, the illumination distribution template and the patrol pictures are fused by linear interpolation. The illumination distribution template is superimposed in different areas to achieve the simulated illumination effect. The application steps are as follows:

[0079] 1. Align the illumination distribution template with the size of the patrol image: Since patrol images have multiple resolutions, in order to be able to be superimposed on the patrol images, the illumination distribution template needs to be scaled to make its size consistent with the patrol image. By comparing the patrol image size and the illumination distribution template size, the illumination distribution template is scaled.

[0080] 2. Due to strong light in the backlit shooting, the camera will automatically shorten the camera shutter, resulting in a dark image overall. In order to simulate the effect of darkening the non-illuminated area when shooting against the light, the overall brightness of the original material needs to be reduced. This embodiment uses exponential brightness adjustment to reduce the brightness of the material. Using exponential brightness adjustment can retain more details in low-brightness areas. The specific exponential brightness adjustment rule is: perform the operation for the RGB value of each pixel on the image: After this operation, a new picture P1 is obtained.

[0081] 3. Apply the illumination distribution template to the new image P1 obtained in the second step. Perform linear interpolation on each pixel on image P1 and the pixel corresponding to the illumination distribution template, so as to achieve the purpose of fusion with the illumination distribution template. The linear interpolation calculation formula is: . Where {R, G, B} is the pixel value of each pixel of the P1 image, and V is the grayscale value of the corresponding position on the illumination distribution template. After this operation, the final image P2 is obtained.

[0082] At present, the industry generally adopts the Yolov5 model as the framework for target detection. However, the generalization ability of Yolov5 relies on the sample size. This embodiment uses Yolox as the framework of the negative sample recognition model. Compared with the Yolov5 model, the Yolox model has better generalization ability. This means that this system relies less on the number of samples than the Yolov5 model and has a faster training speed.

[0083] This embodiment uses LabelImg to label negative sample defects. By conducting analysis and statistics on defects with high misjudgment rates, negative sample targets that need to be identified are obtained. At the same time, in drone photos, the same type of targets have large differences in display details due to the distance of shooting. The anti-vibration hammer farthest from the shooting point can only see the overall outline, while the anti-vibration hammer closest to the shooting point can show more details. In order to avoid the problem of inaccurate model recognition due to such multi-scale differences, this embodiment adopts a multi-dimensional defect labeling method to solve the problem of inaccurate recognition due to excessive scale differences. The specific measures are as follows: Combined with the structural characteristics of the target object, the foreground target object is decomposed into multiple local features for labeling; in addition to the local structure, the overall structure is also covered. Based on the above measures, the generalization ability of the negative sample recognition model prediction can be greatly improved.

[0084] In step S3-2 of this embodiment, after the segmented sub-image is sent to the negative sample recognition model for recognition, the position of the recognized target object is relative to the segmented sub-image, so it needs to be converted into the coordinates of the patrol image.

[0085] At the same time, since there is overlap between the segmented sub-images, the same target will be marked on both sub-images, so target merging is also required.

[0086] In summary, in order to solve the above two problems, this embodiment adopts the following method to merge the recognition target results of each sub-image in the negative sample recognition model:

[0087] Step S3-2-1: Traverse each sub-image after segmentation (q is the total number of sub-images after segmentation). Both call the negative sample recognition model for recognition and obtain the recognition target result set ,in For sub-image The nth recognition target information, e is the number of recognition targets, is the target type of the identified target, Is the recognition target in the sub-image The bounding box information on The definition is ,in The bounding boxes are in the sub-images The minimum value in the X-axis direction, the maximum value in the X-axis direction, the minimum value in the Y-axis direction, and the maximum value in the Y-axis direction.

[0088] Step S3-2-2: Collect each recognition target result The bounding box coordinates of the identified target are converted to the bounding box coordinates relative to the patrol image. The specific steps are: traverse the target recognition result set Each identification target information in , read its bounding box information ,Will Add sub-picture The coordinates of the upper left corner on the patrol image ;Right now .

[0089] Step S3-2-3: Construct a set of final target recognition results ,in To finally identify the target information, is the target type of the final identification prediction target, is the bounding box information of the final recognition target recorded on the patrol image, and k is the number of recognition records.

[0090] Step S3-2-4: Traverse the sub-pictures one by one The recognition target result set .for Each identification target information With each final recognition target information in the set N Make comparisons and determine the target type and Are they the same?

[0091] Step S3-2-5: If they are equal, merge and The bounding box of Get the merged bounding box The merging rules are in It is a record The bounding box parameters in , It is a record The bounding box parameters in ; and the bounding box Update to record of superior.

[0092] Step S3-2-6: If they are not equal, it means that it is a newly discovered prediction target, then record Put it into the set N of final target recognition results, and the set N of final target recognition results is the judgment result of the negative sample recognition model.

[0093] When the system receives the defect judgment results returned by the defect recognition model (external model), the multi-model cascade system calls different negative sample recognition models for different judgment results (such as foreign matter in wires, foreign matter in channels, etc.), conducts comprehensive judgment, and finally forms a comprehensive judgment result. Figure 2 As shown, the multi-model cascade system includes a scheduling module and a comprehensive analysis module, and the defect comprehensive analysis algorithm includes the following steps:

[0094] Step S5-1: First, after receiving the inspection picture and the corresponding defect set M extracted by the defect recognition model, the scheduling module searches one by one according to the defect type in the defect set M to see if there is a corresponding negative sample recognition model. If there is, these defects are extracted separately to form a set A, and the negative sample recognition model is called to perform a secondary judgment on the set A; the part of the defect set M excluding the set A is the set P;

[0095] Step S5-2: The scheduling module calls the negative sample recognition model to judge the set A, and forms a negative sample judgment result set B;

[0096] Step S5-3: After receiving the set A, P, and B, the comprehensive evaluation module performs comprehensive evaluation on the defect. The evaluation process is as follows:

[0097] Step S5-3-1: traverse each defect record a in set A to obtain a bounding box c;

[0098] Step S5-3-2: traverse the negative sample judgment result set B for the bounding box c, and determine whether the defect record in the negative sample judgment result set B intersects with c;

[0099] Step S5-3-3: If they intersect, remove the defect record a from set A;

[0100] Step S5-3-4: Determine whether all elements in set A have been traversed. If not, repeat steps S5-3-1 to S5-3-3 until all elements in set A have been traversed.

[0101] Step S5-3-5: After the traversal is completed, set A and set P are merged to form the final comprehensive analysis result set E.

[0102] To implement the above method, this embodiment provides a multi-model cascade transmission line defect identification false alarm rate optimization system, such as Figure 3 As shown in the figure, it includes three parts: defect recognition model, negative sample recognition model library and multi-model cascade system.

[0103] Defect recognition model: Preliminary identification of defect information in the inspection image, including defect type and defect location;

[0104] Multi-model cascade system: receives inspection images and defect information output by the defect recognition model; calls the corresponding negative sample recognition model from the negative sample recognition model library according to the defect type to conduct comprehensive analysis and judgment, eliminates erroneous defect information, and finally forms defect information after comprehensive analysis and judgment;

[0105] Negative sample recognition model library: includes several different negative sample recognition models. Negative sample recognition models are built for defect types with high misjudgment rates and their corresponding misjudgment scenarios to screen out defects that are misjudged by the defect recognition model.

[0106] The embodiments described above are only descriptions of the preferred methods of the present application, and are not intended to limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements made to the technical solutions of the present application by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present application.

Claims

1. A multi-model cascade transmission line defect identification false alarm rate optimization method, characterized in that: include: Step S1: receiving a patrol picture; Step S2: calling the defect recognition model to perform defect recognition on the inspection image, and extracting the defect type and defect location of each defect; Step S3: Call the corresponding negative sample recognition model according to the defect type to perform negative sample recognition and obtain the negative sample recognition model judgment result; Step S4: Compare the judgment result of the negative sample recognition model with the defect type and defect location extracted by the defect recognition model to determine whether it is a false alarm defect; if it is not a false alarm defect, put the defect into the output defect information; if it is a false alarm defect, remove the defect; Step S5: Traverse all defects in the inspection image and call different negative sample recognition models respectively, conduct comprehensive analysis and judgment based on the output defect information, and finally form a comprehensive analysis and judgment result, including the following steps: Step S5-1: First, after receiving the inspection picture and the corresponding defect set M extracted by the defect recognition model, the scheduling module searches one by one according to the defect type in the defect set M to see if there is a corresponding negative sample recognition model. If there is, these defects are extracted separately to form a set A, and the negative sample recognition model is called to perform a secondary judgment on the set A; the part of the defect set M excluding the set A is the set P; Step S5-2: The scheduling module calls the negative sample recognition model to judge the set A, and forms a negative sample judgment result set B; Step S5-3: After receiving the sets A, P, and B, the comprehensive analysis module conducts comprehensive analysis on the defects. The process is as follows: Step S5-3-1: traverse each defect record a in set A to obtain a bounding box c; Step S5-3-2: traverse the negative sample judgment result set B for the bounding box c, and determine whether the defect record in the negative sample judgment result set B intersects with c; Step S5-3-3: If they intersect, remove the defect record a from set A; Step S5-3-4: Determine whether all elements in set A have been traversed. If not, repeat steps S5-3-1 to S5-3-3 until all elements in set A have been traversed. Step S5-3-5: After the traversal is completed, set A and set P are merged to form the final comprehensive analysis result set E.

2. According to claim 1, a multi-model cascade transmission line defect identification false alarm rate optimization method is characterized in that: The specific process of step S3 is as follows: Step S3-1: Enhance the patrol image according to the patrol features, and then use the enhanced image samples to train the negative sample recognition model; Step S3-2: Realize the recognition of negative samples through the multi-overlapping negative sample recognition model calling strategy; crop the patrol image into several overlapping sub-images, and send the sub-images to the negative sample recognition model for recognition, merge the recognition target results of each sub-image in the negative sample recognition model, and then output the negative sample recognition model judgment result.

3. A multi-model cascade transmission line defect identification false alarm rate optimization method according to claim 2, characterized in that: The enhancement processing includes contrast enhancement and backlight processing.

4. A multi-model cascade transmission line defect identification false alarm rate optimization method according to claim 3, characterized in that: The contrast enhancement process is as follows: first, the image histogram distribution is calculated, and then based on the distribution of the image histogram, the image histogram equalization is performed.

5. A multi-model cascade transmission line defect identification false alarm rate optimization method according to claim 4, characterized in that: The steps of image histogram equalization are as follows: According to the image histogram distribution, calculate the average brightness value; Iterate over each pixel and increase the contrast by increasing the brightness difference between pixels.

6. A multi-model cascade transmission line defect identification false alarm rate optimization method according to claim 3, characterized in that: The backlight processing includes: constructing illumination distribution templates of multiple directions and intensities, and applying the illumination distribution templates to the patrol pictures.

7. A multi-model cascade transmission line defect identification false alarm rate optimization method according to claim 6, characterized in that: For patrol pictures, linear interpolation is used to fuse the illumination distribution template and patrol pictures, and the illumination effect is simulated by superimposing the illumination distribution template in different areas.

8. The method for optimizing the false alarm rate of multi-model cascade transmission line defect identification according to claim 2, characterized in that: In step S3-2, the recognition target results of each sub-image in the negative sample recognition model are merged: Step S3-2-1: Traverse each sub-image after segmentation , q is the total number of sub-images after segmentation; for each sub-image Both call the negative sample recognition model for recognition and obtain the recognition target result set ,in For sub-image The nth recognition target information, e is the number of recognition targets, is the target type of the identified target, Is the recognition target in the sub-image The bounding box information on The definition is ,in The bounding boxes are in the sub-images The minimum value in the X-axis direction, the maximum value in the X-axis direction, the minimum value in the Y-axis direction, and the maximum value in the Y-axis direction; Step S3-2-2: Collect each recognition target result The bounding box coordinates of the identified target are converted to the bounding box coordinates relative to the patrol image. The specific steps are: traverse the identified target result set Each identification target information in , read its bounding box information ,Will Add sub-picture The coordinates of the upper left corner on the patrol image ; Step S3-2-3: Construct a set of final target recognition results ,in To finally identify the target information, is the target type of the final identification prediction target, It is the bounding box information of the final recognition target recorded on the patrol image. k To identify the number of records; Step S3-2-4: Traverse the sub-pictures one by one The recognition target result set ;for Each identification target information With Collection N Each final recognition target information in Make comparisons and determine the target type and Are they the same? Step S3-2-5: If equal, merge and The bounding box of Get the merged bounding box The merging rules are in It is a record The bounding box parameters in , It is a record The bounding box parameters in ; and the bounding box Update to record of superior; Step S3-2-6: If they are not equal, it means that it is a newly discovered prediction target, then record Put it into the collection of final target recognition results N The final target recognition result is the collection of N This is the result of the negative sample recognition model.

9. A multi-model cascade transmission line defect identification false alarm rate optimization system, used to implement the multi-model cascade transmission line defect identification false alarm rate optimization method according to any one of claims 1 to 8, characterized in that: It includes three parts: defect recognition model, negative sample recognition model library and multi-model cascade system: Defect recognition model: Preliminary identification of defect information in the inspection image, including defect type and defect location; Multi-model cascade system: receives inspection images and defect information output by the defect recognition model; calls the corresponding negative sample recognition model from the negative sample recognition model library according to the defect type to conduct comprehensive analysis and judgment, eliminates erroneous defect information, and finally forms defect information after comprehensive analysis and judgment; Negative sample recognition model library: includes several different negative sample recognition models, which are used to filter out defects that are misjudged by the defect recognition model.

Images