Disk-shaped insulator angle adaptive target detection method and system based on structural characteristics

Abstract

The invention relates to a structural feature-based disc-shaped insulator angle adaptive target detection method and system. The method comprises the following steps of collecting a disc-shaped insulator inspection picture, and performing preprocessing and manual labeling; parameterizing a rotating target frame of the disc-shaped insulator by using the annotation data and forming a data set; establishing a disc-shaped insulator angle adaptive target detection network based on structural characteristics; training a target detection network by using the training set, evaluating a training result by using an improved rotating bounding box IoU, and solidifying a disk-shaped insulator angle self-adaptive target detection network model; and detecting a to-be-detected insulator picture by using the trained target detection network to obtain a disc-shaped insulator rotation target detection frame. According to the method, the background contained in the inclined distribution insulator detection frame can be effectively reduced, the detection distinction degree of the densely-arranged insulators is increased, the structural direction distribution of the insulators is reflected, the sensitivity problem of angle variable prediction is avoided, and the precision and speed of rotating target detection of the disc-shaped insulators are improved.

Description

Technical Field

[0001] The present application relates to the technical field of image target detection, and more specifically, to a method and system for disc-shaped insulator angle adaptive target detection based on structural features. Background Art

[0002] Intelligent inspection is one of the main development directions of power line inspection today. Disc insulators are important components of power lines and are also the focus of power inspection. Intelligent recognition and detection of insulators through inspection images is an important basis for many intelligent inspection tasks, such as refined detection of insulator defects and auxiliary intelligent inspection and photography of key power components. However, the current insulator target detection algorithm mainly identifies the target with a horizontal axial bounding box. For objects with a large aspect ratio such as insulators, when the target is in an inclined distribution, the axial bounding box contains a large amount of background, and it is difficult to distinguish densely arranged targets, which has an adverse effect on subsequent intelligent tasks.

[0003] Using a rotating bounding box with angle information to mark insulators, namely the rotating target detection method, can effectively solve the aforementioned problems. However, the current rotating target detection method is mainly designed for remote sensing scenarios, and the algorithm needs to add a large number of complex designs, such as rotating attention mechanism, rotating candidate box generation, etc., to cope with the special needs of the bounding box rotation angle prediction task, which has a serious impact on the algorithm operation efficiency. Some methods are sensitive to angle variables, resulting in problems such as degradation of detection performance, and have certain limitations in their application in power inspection scenarios.

[0004] Prior art such as the Chinese patent with patent number "CN116704211A" is disclosed. This method uses the structural characteristics of each typical component of the transmission line to define its structural key points, and then formulates a corresponding labeling scheme, and uses the innovative concept of structural key points to achieve target and defect identification of typical components of the transmission line in a complex environment, thereby reducing the risk of safety hazards caused by defects in the transmission line body and its components. The present invention includes: designing and defining the structural key points of each typical component of the transmission line according to the inspection guidelines; formulating the corresponding labeling criteria and labeling result output data of each typical component of the transmission line according to the defined structural key points; designing a target detection method for typical components of the transmission line based on key point detection, thereby effectively improving the efficiency and accuracy of power grid inspection.

[0005] The problems existing in the above-mentioned prior art are: target recognition of typical components of transmission lines is performed only through key points and rectangular frames. When the target is in an oblique distribution, the rectangular bounding box contains a large amount of background, and it is difficult to distinguish densely arranged targets, which is not conducive to the analysis and processing of subsequent intelligent tasks. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes a disc insulator angle adaptive target detection method and system based on structural features.

[0007] The technical solution of the present invention is as follows:

[0008] The present invention proposes a disc insulator angle adaptive target detection method based on structural features, comprising the following steps:

[0009] Step S1, collecting inspection pictures of disc insulators, and pre-processing and manually labeling the collected inspection pictures;

[0010] Step S2, parameterizing the disc insulator rotation target frame using the labeled data and forming a data set, the data set including a training set and a test set;

[0011] Step S3, establishing a disc insulator angle adaptive target detection network based on structural features;

[0012] Step S4, using the training set to train the disc insulator angle adaptive target detection network, using the improved rotation bounding box IoU to evaluate the training results and solidifying the disc insulator angle adaptive target detection network model;

[0013] Step S5, using the trained disc insulator angle adaptive target detection network to detect the disc insulator image to be detected, and obtaining a disc insulator rotating target detection frame reflecting the direction distribution of the insulator structure.

[0014] As a preferred embodiment, the collected inspection pictures are pre-processed and manually annotated, wherein the pre-processing includes expanding the four sides of the picture respectively and filling them with black; the annotation content includes: taking the direction of the insulator iron cap as the positive direction of the insulator rotation target, marking the insulator rotation target frame and the center point of the middle insulator piece of the visible part of the insulator string image.

[0015] As a preferred embodiment, the disk insulator rotation target frame is re-parameterized using the annotation data to form a data set, wherein the parameterization content includes: for an incompletely displayed disk insulator string image, the pixel coordinates (x s ,y s ), the size of the insulator rotation bounding box (w, h), the offset of the center point of the rotation bounding box relative to the pixel coordinates of the center point of the middle insulator piece (Δx c ,Δy c ), the offset of the pixel coordinates of the midpoint of the positive terminal of the rotation bounding box relative to the center point of the middle insulator piece (Δx k ,Δx k ).

[0016] As a preferred implementation, the disc insulator angle adaptive target detection network based on structural features is established; the detection network includes an image preprocessing layer, a feature extraction backbone network, a plurality of output heads and an output decoder, wherein:

[0017] Image preprocessing layer, which expands the image boundary and fills the color of the input image, adjusts the image resolution, and fills the insufficient part with black;

[0018] The feature extraction backbone network uses a fast-trained residual network to extract features of the input image;

[0019] The output head includes at least one thermal map output head for detecting the center point of the middle insulator piece of the visible part of the disc-shaped insulator; one size parameter output head for predicting the size of the rotating bounding box; and one offset parameter output head for predicting the offset between the center point of the middle insulator piece and the center point of the rotating bounding box, and the offset between the center point of the middle insulator piece and the midpoint of the terminal in the positive direction of the rotating bounding box.

[0020] The output decoder receives the output data of the output head and obtains the size of the rotated bounding box from (w,h); s ,y s ) and (Δx c ,Δy c ) predicts the value to obtain the coordinates of the center point of the disc insulator rotation bounding box; by (x s ,y s ) and (Δx k ,Δx k ) prediction value to obtain the coordinates of the midpoint of the positive terminal of the rotating bounding box; the center point of the disc insulator rotation detection box and the midpoint vector of the positive terminal of the rotating bounding box are used to obtain the rotation angle θ of the rotating bounding box, and finally the five parameters (x c ,y c ,w,h,θ), and convert them into the coordinates of the four vertices of the rotated bounding box for the visualization output of the detection results.

[0021] As a preferred implementation, the method of using a data set to train a disc insulator angle adaptive target detection network, using an improved rotation bounding box IoU to evaluate the training results and solidifying the disc insulator angle adaptive target detection network model comprises the following steps:

[0022] The image preprocessing layer of the shielding detection network adjusts the input image resolution through the training interface, fills the insufficient part with black, and recalculates the annotation parameters of the insulator rotation target bounding box based on the resolution adjustment ratio and the downsampling multiple;

[0023] The loss function is constructed through the output data of the output head to train the detection network; the heat map output head uses the FOCAL loss function, and the size parameter output head and the offset parameter output head use the Smooth L1 loss function;

[0024] The average accuracy of the adaptive target detection network detected by the test set is used to save the optimal network model parameters; the average accuracy of the adaptive target detection network is detected, and the improved DOTA public dataset evaluation method is used. In the DOTA evaluation method, a rotation bounding box direction angle judgment threshold of 90° is added to filter out the high IoU detection results with opposite heads and tails of insulators.

[0025] As a preferred implementation, the loss function is constructed through the output data of the output head; the total training loss function is obtained through the heat map output head, the size parameter output head and the offset parameter output head. The specific calculation formula is as follows:

[0026] L total =λ ht L ht +λ s L s +λ o L o ;

[0027] In the formula, λ ht , s and λ o are the training loss weights of the corresponding output heads respectively.

[0028] On the other hand, the present invention also provides a disc insulator angle adaptive target detection system based on structural features, comprising:

[0029] The data annotation module is used to collect inspection pictures of disc insulators, and to pre-process and manually annotate the collected inspection pictures;

[0030] The dataset construction module uses the labeled data to parameterize the disc insulator rotation target box and construct a dataset, which includes a training set and a test set;

[0031] Model building module, building a disc insulator angle adaptive target detection network based on structural features;

[0032] Model training and evaluation module: Use the training set to train the disc insulator angle adaptive target detection network, use the improved rotation bounding box IoU to evaluate the training results and solidify the disc insulator angle adaptive target detection network model;

[0033] The detection and recognition module uses the trained disc insulator angle adaptive target detection network to detect the disc insulator image to be detected, and obtains the disc insulator rotating target detection frame that reflects the direction distribution of the insulator structure.

[0034] As a preferred embodiment, the data annotation module preprocesses and manually annotates the collected inspection images, wherein the preprocessing includes expanding the four sides of the image respectively and filling them with black; the annotation content includes: taking the direction of the insulator iron cap as the positive direction of the insulator rotation target, annotating the insulator rotation target frame and the center point of the middle insulator piece of the visible part of the insulator string image.

[0035] On the other hand, the present invention further provides an electronic device having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for disc insulator angle adaptive target detection based on structural features as described in any embodiment of the present invention is implemented.

[0036] On the other hand, the present invention also provides a computer-readable medium for storing one or more programs, which, when executed by the one or more processors, enables the one or more processors to implement a disc insulator angle adaptive target detection method based on structural features as described in any embodiment of the present invention.

[0037] The present invention has the following beneficial effects:

[0038] 1. The present invention uses a rotating bounding box to mark insulators. Compared with the traditional insulator axial target detection algorithm, when the disc-shaped insulators are obliquely distributed, the background noise contained is less, and the distinction between densely arranged insulators is higher, which is more conducive to the subsequent tasks such as refined detection of insulator defects, auxiliary intelligent inspection and photography of key power components, and visual spatial positioning of insulators; compared with the semantic segmentation algorithm, the present invention does not require pixel-level annotation, the data annotation cost is lower, and it is more conducive to engineering applications.

[0039] 2. The method for defining the rotation angle of the disc-shaped insulator enclosing frame based on structural features proposed in the present invention is better able to reflect the actual distribution direction of the insulator, and is more conducive to the segmentation, identification and accurate warning of defect locations of insulator-related key inspection areas such as insulator crossarm hanging points and insulator conductor hanging points, compared to the long side definition method of the rotation angle used in the traditional rotating target detection algorithm, that is, defining the rotation angle of the rectangular frame by the angle between the long side of the rotating rectangular frame and the horizontal x-axis of the image.

[0040] 3. The structural feature-based angle-adaptive target detection method for disc insulators proposed in the present invention calculates the rotation angle of the insulator bounding box by predicting the relative positions of the key structural features of the insulator components. Compared with the traditional rotating target detection that directly predicts the rotation angle variable, it avoids the problem of rotation angle sensitivity in the rotating target detection method, and does not require module designs for rotating target detection tasks such as rotating receptive field and rotating bounding box suggestion generation. It improves the detection accuracy and operation efficiency of insulator rotating target detection, and has better adaptability to task scenarios with high requirements on algorithm real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0042] Figure 1 It is the work flow chart of the present invention.

[0043] Figure 2 It is a schematic diagram of the disc insulator data labeling method of the present invention.

[0044] Figure 3 This is a schematic diagram of the detection principle of the present invention.

[0045] In the figure: 1. The center point of the middle insulator piece of the visible part of the insulator string image; 2. The positive direction of the insulator rotation target; 3. The midpoint of the positive direction terminal of the rotation bounding box calculated using the predicted offset; 4. The direction of the rotation bounding box calculated using the center point of the rotation bounding box and the midpoint of the positive direction terminal; 5. The center point of the middle insulator piece of the visible part of the insulator string is detected; 6. The center point of the insulator rotation bounding box calculated using the predicted offset. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] It should be understood that the step numbers used in this document are only for convenience of description and are not intended to limit the order in which the steps are executed.

[0048] It should be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0049] The terms “include” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0050] The term "and / or" means and includes any and all possible combinations of one or more of the associated listed items.

[0051] Embodiment 1:

[0052] In order to make the purpose, technical solutions and advantages of the present invention more clear, the following will be combined with the specific embodiments of the present application and refer to the attached Figure 1 , clearly and completely describe the technical solution of the present invention.

[0053] The purpose of the present invention is to provide an angle-adaptive target detection method for disc insulators, so as to overcome the problems that the axial bounding box of the traditional insulator target detection algorithm contains a large amount of background and it is difficult to distinguish densely arranged targets, and the existing rotating target detection methods have poor algorithm operation efficiency in power scene applications and some methods are sensitive to angle variables, so as to realize rapid rotating target detection of disc insulators.

[0054] To solve the problems in the prior art, the present invention provides a disc insulator angle adaptive target detection method based on structural features, comprising the following steps:

[0055] Step S1: See Figure 2 As shown in the figure, the inspection pictures of disc insulators are collected, and the collected inspection pictures are preprocessed and manually annotated; the preprocessing includes: selecting the inspection pictures of disc insulators, expanding the four sides of the pictures by 15% respectively, and filling them with black. The manual annotation includes: taking the direction of the insulator iron cap as the positive direction of the insulator rotation target, using the annotation tool to respectively annotate the rotation bounding box of the insulator and the center point of the middle insulator piece of the visible part of the insulator string image; for the incompletely displayed insulator string, the farthest visible end of the central axis of its components is the terminal of the bounding box. The generated annotation data are the pixel coordinates of the center point of the rotating bounding box (x c ,y c), the pixel size width and length of the bounding box (w, h), the clockwise rotation angle θ between the positive direction of the bounding box and the negative direction of the image coordinate y axis (range [0, 360°)), and the pixel coordinates of the center point of the middle insulator piece of the visible part of the insulator string image (x s ,y s ).

[0056] Step S2: Parameterize the marked insulator rotation target frame as (x s ,y s ,w,h,Δx c ,Δy c ,Δx k ,Δy k ), where (x s ,y s ) represents the pixel coordinates of the center point of the middle insulator piece in the visible part of the insulator string image, which is directly obtained from the annotation data; (w,h) are the width and height of the insulator rotation bounding box, which are directly obtained from the annotation data; (Δx c ,Δy c ) represents the pixel coordinates of the center point of the middle insulator piece in the visible part of the insulator string image (x s ,y s ) and the pixel coordinates of the center point of the rotation bounding box (x c ,y c ) pixel offset, given by (x s ,y s )-(x c ,y c ) is calculated; (Δx k ,Δx k ) represents the pixel coordinates of the center point of the middle insulator piece in the visible part of the insulator string image (x s ,y s ) and the pixel offset of the midpoint pixel coordinates of the terminal in the positive direction of the rotation bounding box, which is calculated by the following formula:

[0057]

[0058] The rotation target box parameters of all disc insulator instances in the image are written into the same json file, named after the image file name after boundary expansion. The image data after boundary expansion and the image annotation data constitute the data set; the data set includes a training set and a test set.

[0059] Step S3: Establish a disc insulator angle adaptive target detection network based on structural features, including an image preprocessing layer, a feature extraction backbone network, three output heads and an output decoder. The image preprocessing layer is responsible for performing boundary extension and filling processing on the input image as described in step S1. The three output heads include: a heat map output head, which is responsible for detecting the center point of the middle insulator piece of the visible part of the insulator; a size parameter output head is responsible for predicting the size of the rotation bounding box; and an offset parameter output head is responsible for predicting the offset between the center point of the middle insulator piece and the center point of the rotation bounding box, and the offset between the center point of the middle insulator piece and the midpoint of the positive terminal of the rotation bounding box. The output decoder is responsible for integrating the output data and decoding it into the final insulator rotation target box.

[0060] Step S301: Image preprocessing layer, the input image is extended and filled with color according to the method described in step S1, and the image resolution is adjusted to m×m pixels, and the insufficient part is filled with black. m is a constant parameter set manually, which is set to 512.

[0061] Step S302: The feature extraction backbone network uses Resnet-101, the image input resolution is unified as m×m pixels, the downsampling multiple is set to n, and n is a manually set constant parameter set to 4; the heat map output head is responsible for the detection of the center point of the middle insulator piece of the visible part of the insulator, which consists of a 3×3 convolution layer and a 1×1 convolution layer, and the output size is The pixel value range is [0,1]; the size parameter output head and the offset parameter output head are composed of a 3×3 convolution layer and a 1×1 convolution layer, respectively, and the output sizes are and

[0062] Step S303: See Figure 3 As shown, the output decoder receives the output data of the three output heads, selects the first k thermal points with the largest thermal values ​​in the thermal map, and sets the pixel coordinates of the thermal points that meet the requirements as The corresponding size parameter prediction value is The predicted value of the offset parameter is The center point of the rotation bounding box of the m×m pixel input image is calculated by the following formula: Rotate the bounding box in the positive direction to the terminal midpoint and the bounding box size

[0063]

[0064] Then the unit direction vector of the insulator bounding box of the m×m pixel input image is calculated by the following formula:

[0065]

[0066] Where ck represents the vector from the center point of the rotating bounding box to the midpoint of the positive terminal of the rotating bounding box; u r and v r Respectively The directional components of the x and y axes of the image.

[0067] The rotation angle parameter of the rotation bounding box of the m×m pixel input image is calculated by the following formula:

[0068]

[0069] Thus, the five parameters of the insulator bounding box of the m×m pixel input image are obtained And convert it into the pixel coordinates of the four vertices of the rotated bounding box of the m×m pixel input image.

[0070] According to the structural features of the disc-shaped insulator, the center point of the rotating bounding box is close to the center point of the middle insulator piece of the insulator string with more significant visual features; the midpoint of the terminal in the positive direction of the rotating bounding box is close to the position of the top iron cap of the insulator with more significant visual features, and the top iron cap of the insulator and the center point of the middle insulator piece of the insulator string have more significant image connection features. The method proposed in the present invention utilizes the above-mentioned disc-shaped insulator structural features to model the insulator rotation target detection task as an insulator structural feature detection and corresponding offset prediction problem, thereby avoiding the angle variable sensitivity problem and special module design of the rotation detection task existing in the traditional rotation target detection method, and improving the detection accuracy and operation efficiency of the insulator rotation target detection.

[0071] Step S4: Input the training set constructed in step S2 into the disc insulator angle adaptive target detection network described in step S3, train the detection network, evaluate the training results through the test set and use the improved rotation bounding box IoU to solidify the detection network model.

[0072] Step S4.1: The image preprocessing layer of the mask detection network is used to adjust the input image resolution to m×m pixels by the training interface, and the insufficient part is filled with black. s ,y s ,w,h,Δx c ,Δy c ,Δx k ,Δy k ) is recalculated according to the resolution adjustment ratio and divided by the sampling multiple n to generate The true value of the rotation target bounding box under size Among them, the true value of the center point of the middle insulator piece of the visible part of the insulator It is necessary to round down and use the Gaussian kernel to draw the true value of the heat map:

[0073]

[0074] Among them, x and y are the pixel coordinates of the heat map; σ p is the Gaussian radius, which can be calculated as follows:

[0075]

[0076] Where ol is the IoU threshold of the predicted box that is expected to be retained, which is an artificially set constant parameter and is set to 0.7; r1, r2 and r3 are the deviation radii of the box vertices in the following three cases: 1. The predicted box completely wraps the real box; 2. The predicted box is completely wrapped by the real box; 3. The predicted box and the real box partially overlap.

[0077] Step S4.2: Construct a loss function to train the detection network through the output data of the three output heads. The heat map output head uses FOCAL loss as the training loss function, and its calculation formula is as follows:

[0078]

[0079] Where α and β are the loss function hyperparameters, and N is the number of targets.

[0080] Both the size parameter output head and the offset parameter output head use Smooth L1 to construct the training loss function:

[0081]

[0082] Where i represents the i-th target instance.

[0083] The total training loss function calculation formula is as follows:

[0084] L total =λ ht L ht +λ s L s +λ o L o ;

[0085] In the formula, λ ht , s and λ o are the training loss weights of the corresponding output heads respectively.

[0086] When training the network, set the training loss weight λ of each output head ht =1,λ s = 0.6 and λ o =0.8; initial learning rate 7×10 -3, and decreased to 7×10 at 200 and 230 training rounds respectively. -5 and 7×10 -6 ; The Adam optimizer is used to optimize the overall objective, and the first-order momentum and second-order momentum are 0.9 and 0.999 respectively.

[0087] Step S4.3: Detect AP based on the validation set and save the best network model parameters. The AP is detected using an improved DOTA public dataset evaluation method. In the DOTA evaluation method, a 90° rotation bounding box direction angle determination threshold is added to filter out high IoU detection results where the insulator head and tail are opposite.

[0088] Step S5: In the detection network application stage, the trained disc insulator angle adaptive target detection network is used to activate the image preprocessing layer. The detection image is first extended and filled by the image preprocessing layer, and then the detection network is used to obtain the parameters of the insulator rotation target bounding box (x s ,y s ,w,h,Δx c ,Δy c ,Δx k ,Δy k ), and finally, the detection result of the decoder insulator rotation bounding box is output, and the detection result is visualized as follows Figure 2 The rotating bounding box with the insulator positive direction mark is shown, and the center point mark of the middle insulator piece of the visible part of the insulator string image is not marked.

[0089] Embodiment 2:

[0090] This embodiment provides a disc insulator angle adaptive target detection system based on structural features, including:

[0091] The data annotation module is used to collect inspection pictures of disc insulators, and to pre-process and manually annotate the collected inspection pictures;

[0092] The dataset construction module uses the labeled data to parameterize the disc insulator rotation target box and construct a dataset, which includes a training set and a test set;

[0093] Model building module, building a disc insulator angle adaptive target detection network based on structural features;

[0094] Model training and evaluation module: Use the training set to train the disc insulator angle adaptive target detection network, use the improved rotation bounding box IoU to evaluate the training results and solidify the disc insulator angle adaptive target detection network model;

[0095] The detection and recognition module uses the trained disc insulator angle adaptive target detection network to detect the disc insulator image to be detected, and obtains the disc insulator rotating target detection frame that reflects the direction distribution of the insulator structure.

[0096] Embodiment three:

[0097] This embodiment provides an electronic device having a computer program stored thereon. When the computer program is executed by a processor, the method for disc insulator angle adaptive target detection based on structural features as described in any embodiment of the present invention is implemented.

[0098] Embodiment 4:

[0099] This embodiment provides a computer-readable medium for storing one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement a disc insulator angle adaptive target detection method based on structural features as described in any embodiment of the present invention.

[0100] In the embodiments of the present application, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. Among them, A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c can be represented by: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, c can be single or multiple.

[0101] Those of ordinary skill in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented in a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0102] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0103] In several embodiments provided in the present application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), disk or optical disk, and other media that can store program codes.

[0104] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A disc insulator angle adaptive target detection method based on structural features, characterized in that: The following steps are involved: Step S1, collecting inspection pictures of disc insulators, pre-processing and manually marking the collected inspection pictures; Step S2, parameterizing the disc insulator rotation target frame using the labeled data and forming a data set, the data set including a training set and a test set; Step S3, establishing a disc insulator angle adaptive target detection network based on structural features; Step S4, using the training set to train the disc insulator angle adaptive target detection network, using the improved rotation bounding box IoU to evaluate the training results and solidifying the disc insulator angle adaptive target detection network model; Step S5, using the trained disc insulator angle adaptive target detection network to detect the disc insulator image to be detected, and obtaining a disc insulator rotating target detection frame reflecting the direction distribution of the insulator structure.

2. The disc insulator angle adaptive target detection method based on structural features according to claim 1 is characterized in that: The collected inspection pictures are preprocessed and manually annotated, wherein the preprocessing includes expanding the four sides of the picture respectively and filling them with black; the annotation content includes: taking the direction of the insulator iron cap as the positive direction of the insulator rotation target, marking the insulator rotation target frame and the center point of the middle insulator piece of the visible part of the insulator string image.

3. The disc insulator angle adaptive target detection method based on structural features according to claim 1, characterized in that: The disk insulator rotation target frame is re-parameterized by using the annotation data to form a data set, wherein the parameterization content includes: for the incompletely displayed disk insulator string image, the pixel coordinates (x s ,y s ), the size of the insulator rotation bounding box (w, h), the offset of the center point of the rotation bounding box relative to the pixel coordinates of the center point of the middle insulator piece (Δx c ,Δy c ), the offset of the pixel coordinates of the midpoint of the positive terminal of the rotation bounding box relative to the center point of the middle insulator piece (Δx k ,Δx k ).

4. The method for disc insulator angle adaptive target detection based on structural features according to claim 1, characterized in that: The disc insulator angle adaptive target detection network based on structural features is established; the detection network includes an image preprocessing layer, a feature extraction backbone network, a plurality of output heads and an output decoder, wherein: Image preprocessing layer, which expands the image boundary and fills the color of the input image, adjusts the image resolution, and fills the insufficient part with black; The feature extraction backbone network uses a fast-trained residual network to extract features of the input image; The output head includes at least one thermal map output head for detecting the center point of the middle insulator piece of the visible part of the disc-shaped insulator; one size parameter output head for predicting the size of the rotating bounding box; and one offset parameter output head for predicting the offset between the center point of the middle insulator piece and the center point of the rotating bounding box, and the offset between the center point of the middle insulator piece and the midpoint of the terminal in the positive direction of the rotating bounding box. The output decoder receives the output data of the output head and obtains the size of the rotated bounding box from (w,h); s ,y s ) and (Δx c ,Δy c ) predicts the value to obtain the coordinates of the center point of the disc insulator rotation bounding box; by (x s ,y s ) and (Δx k ,Δx k ) prediction value to obtain the coordinates of the midpoint of the positive terminal of the rotating bounding box; the center point of the disc insulator rotation detection frame and the midpoint vector of the positive terminal of the rotating bounding box are used to obtain the rotation angle θ of the rotating bounding box, and finally the five parameters (x c ,y c ,w,h,θ), and convert them into the coordinates of the four vertices of the rotated bounding box for the visualization output of the detection results.

5. The method for disc insulator angle adaptive target detection based on structural features according to claim 1, characterized in that: The method of using a data set to train a disc insulator angle adaptive target detection network, using an improved rotation bounding box IoU to evaluate the training results and solidifying the disc insulator angle adaptive target detection network model comprises the following steps: The image preprocessing layer of the shielding detection network adjusts the input image resolution through the training interface, fills the insufficient part with black, and recalculates the annotation parameters of the insulator rotation target bounding box based on the resolution adjustment ratio and the downsampling multiple; The loss function is constructed through the output data of the output head to train the detection network; the heat map output head uses the FOCAL loss function, and the size parameter output head and the offset parameter output head use the Smooth L1 loss function; The average accuracy of the adaptive target detection network detected by the test set is used to save the optimal network model parameters; the average accuracy of the adaptive target detection network is detected, and the improved DOTA public dataset evaluation method is used. In the DOTA evaluation method, a rotation bounding box direction angle judgment threshold of 90° is added to filter out the high IoU detection results with opposite heads and tails of insulators.

6. The method for disc insulator angle adaptive target detection based on structural features according to claim 5, characterized in that: The loss function is constructed through the output data of the output head; the total training loss function is obtained through the heat map output head, the size parameter output head and the offset parameter output head. The specific calculation formula is as follows: L total =λ ht L ht +λ s L s +λ o L o ; In the formula, λ ht , s and λ o are the training loss weights of the corresponding output heads respectively.

7. A disc insulator angle adaptive target detection system based on structural features, characterized in that: include: The data annotation module is used to collect inspection pictures of disc insulators, and to pre-process and manually annotate the collected inspection pictures; The dataset construction module uses the labeled data to parameterize the disc insulator rotation target box and construct a dataset, which includes a training set and a test set; Model building module, building a disc insulator angle adaptive target detection network based on structural features; Model training and evaluation module: Use the training set to train the disc insulator angle adaptive target detection network, use the improved rotation bounding box IoU to evaluate the training results and solidify the disc insulator angle adaptive target detection network model; The detection and recognition module uses the trained disc insulator angle adaptive target detection network to detect the disc insulator image to be detected, and obtains the disc insulator rotating target detection frame that reflects the direction distribution of the insulator structure.

8. The disc insulator angle adaptive target detection system based on structural features according to claim 7, characterized in that: The collected inspection pictures are preprocessed and manually annotated, wherein the preprocessing includes expanding the four sides of the picture respectively and filling them with black; the annotation content includes: taking the direction of the insulator iron cap as the positive direction of the insulator rotation target, marking the insulator rotation target frame and the center point of the middle insulator piece of the visible part of the insulator string image.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the method for disc insulator angle adaptive target detection based on structural features as described in any one of claims 1 to 6 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for disc insulator angle adaptive target detection based on structural features as described in any one of claims 1 to 6 is implemented.

Images