Power transmission inspection small sample scene identification method based on selective search algorithm

Abstract

The invention discloses a power transmission inspection small sample scene recognition method based on a selective search algorithm, and the method comprises the following steps: S1, collecting an original image of power transmission line inspection, and generating a candidate region according to the original image through the selective search algorithm, extracting a corresponding candidate region from the original image according to the coordinate information of the candidate region as a training sample; s2, establishing a DERT model, performing self-supervised learning pre-training on the DERT model by using a candidate region generated by a selective search algorithm, and performing fine-tuning training on the DERT model through a small number of labeled samples on the basis of self-supervised learning to obtain a trained DERT model; and S3, acquiring a to-be-detected power transmission line inspection image in real time, and performing power transmission line inspection target detection by using the trained DERT model. According to the method, the problems of data scarcity, annotation difficulty and difficulty in efficient detection in electric power inspection are solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power systems, and in particular relates to a small sample scene recognition method for power transmission inspection based on a selection search algorithm. Background Art

[0002] Deep learning is a data-driven technology. The rapid development of deep learning technology in the past 10 years is mainly due to the establishment of large-scale data sets. The development of target detection algorithms is also inseparable from the contribution of data sets. Data sets are not only a common basis for measuring and comparing the performance of target detection algorithms, but also a powerful driving force for the development of target detection. Since power inspection data sets involve some confidential or sensitive information, there is currently no public power inspection data set.

[0003] On the other hand, the premise of deep learning is based on a large amount of labeled data. In the field of power, the labeling of data sets still requires relevant professionals to spend a lot of time and energy to label. Nguyen et al. created a medium-sized power inspection data set by manually labeling images from scratch with 30,000 images with 54 classes. The average labeling speed is 40 images per hour, which makes the construction of power inspection data sets more difficult. Therefore, most research teams conduct research through small data sets they have established. Therefore, it is necessary to carry out research on key information extraction technology for small sample scenarios of power transmission inspection. By studying the power vision large model construction technology for open source large-scale target detection general data sets and the large model efficient fine-tuning method for small sample data of power transmission inspection scenarios, image information extraction and model pre-training can be realized, and a large amount of image prior knowledge can be used to improve the model accuracy in small sample scenarios, so as to realize efficient and low-cost transmission inspection model development, significantly improve the defect detection capability, reduce the intensity of defect detection work, and ensure the safe and stable operation of the transmission system. At the same time, self-supervised learning methods have achieved great success in the fields of natural language processing and image classification, but these methods cannot be directly transferred to target detection technology. Therefore, there is an urgent need for a self-supervised learning method suitable for target detection.

[0004] The Chinese patent with the publication number CN113674225A discloses a method for detecting faults of power equipment based on convolutional neural networks, including constructing a fault type database to train the convolutional neural network for fault recognition; after enhancing the brightness, color, contrast and sharpness of the image to be detected, filtering the detection image with a selection search algorithm with a filter to screen out the candidate area most likely to have a fault; importing the candidate area into the trained convolutional neural network, and obtaining the fault type of the area to be detected and the relative position and size of each fault according to the training rules, wherein the fault types include blurred dial, damaged shell, oily ground, damaged respirator, and abnormal door closure. Although the algorithm described in the invention can identify power equipment faults, it mainly targets surface defects in power equipment fault detection, while the faults of transmission lines are often more complex and have various manifestations, such as line breakage, wear, etc. These problems are difficult to accurately identify through simple image enhancement and convolution feature extraction. Summary of the invention

[0005] The present invention provides a small sample scene recognition method for power transmission inspection based on a selective search algorithm, aiming to solve the problems of data scarcity, labeling difficulty and difficulty in efficient detection in power inspection.

[0006] In order to solve the above technical problems, the present invention provides a method for identifying small sample scenarios of power transmission inspection based on a selection search algorithm, comprising the following steps:

[0007] S1: Collect the original image of the transmission line inspection, generate candidate areas according to the original image by selecting the search algorithm, and extract the corresponding candidate areas from the original image as training samples according to the coordinate information of the candidate areas.

[0008] S2: Establish a DERT model, use the candidate regions generated by the selection search algorithm to perform self-supervised learning pre-training on the DERT model, and fine-tune the DERT model with a small number of labeled samples based on self-supervised learning to obtain a trained DERT model.

[0009] S3: Collect the inspection images of the transmission line to be inspected in real time, and use the trained DERT model to perform transmission line inspection target detection.

[0010] Preferably, the step S1 generates a candidate frame according to the original image by using a search algorithm as follows:

[0011] S11: Perform superpixel segmentation on the original image to divide the original image into multiple regions.

[0012] S12: For each pair of adjacent regions, calculate the total similarity of color, texture, and size similarity between the regions.

[0013] S13: Merge adjacent regions according to similarity to generate new candidate regions.

[0014] S14: Repeat the merging process until the number of generated candidate regions reaches a preset target number, then stop merging and output all generated candidate regions.

[0015] Preferably, the total similarity of the color, texture and size similarities between the calculated regions in step S12 is expressed as:

[0016] S(R i , R j )=ω1S color (R i , R j )+ω2S texture (R i , R j )+ω3S size (R i , R j )

[0017] In the formula, S(R i ,R j ) is the total similarity; R i , R j For each pair of adjacent regions; S color (R i ,R j ) is a similarity measure based on color features; S texture (R i ,R j ) is a similarity measure based on texture features; S size (R i ,R j ) is the similarity measure based on region size; ω1, ω2, ω3 are weight coefficients, the contribution of each feature to the total similarity.

[0018] The calculation formula for color similarity is:

[0019]

[0020] In the formula, For region R i The color mean of For region R j The color mean.

[0021] The calculation formula of texture similarity is as follows:

[0022]

[0023] In the formula, Region R i , Rj The value of the kth bucket in the texture histogram.

[0024] The calculation formula for size similarity is:

[0025]

[0026] In the formula, |R i |、|R j |areas R i , R j area.

[0027] Preferably, in step S2, the improved DERT model is pre-trained by self-supervised learning using the candidate region images generated by the selection search algorithm, and the DERT model is fine-tuned by a small amount of labeled samples based on the self-supervised learning. Specifically, the following steps are performed:

[0028] S21: Divide the candidate region into positive and negative samples according to the characteristics of the candidate region, and perform standardization on each candidate region.

[0029] S22: Design self-supervised learning tasks including region reconstruction, target inference, and contrastive learning, and add them to the DERT model.

[0030] S23: The standardized candidate regions are input into the DERT model for training. The DERT model learns based on the self-supervised task and gradually adjusts the parameters so that the DERT model can capture the target information in the candidate regions.

[0031] S24: Input the labeled samples into the DERT model that has been pre-trained through self-supervised learning, adopt a small batch learning strategy, and use a small amount of labeled data to fine-tune the DERT model.

[0032] On the other hand, the present invention provides a power transmission inspection small sample scene recognition system based on a selective search algorithm, including an image acquisition and candidate area generation module, a DERT model training and fine-tuning module, and a target detection module.

[0033] The image acquisition and candidate region generation module is used to acquire the original image of the transmission line inspection, generate the candidate region according to the original image by selecting the search algorithm, and extract the corresponding candidate region from the original image as the training sample according to the coordinate information of the candidate region.

[0034] The DERT model training and fine-tuning module is used to establish the DERT model. The candidate regions generated by the selection search algorithm are used to pre-train the DERT model through self-supervised learning. Based on the self-supervised learning, the DERT model is fine-tuned through a small number of labeled samples to obtain the trained DERT model.

[0035] The target detection module is used to collect the inspection images of the transmission line to be inspected in real time, and use the trained DERT model to perform transmission line inspection target detection.

[0036] Preferably, the image acquisition and candidate region generation module generates a candidate frame according to the original image by using a search algorithm as follows:

[0037] Perform superpixel segmentation on the original image and divide the original image into multiple regions.

[0038] For each pair of adjacent regions, the total similarity of color, texture, and size similarity between the regions is calculated.

[0039] Adjacent regions are merged according to their similarity to generate new candidate regions.

[0040] The merging process is repeated until the number of generated candidate regions reaches the preset target number, then the merging is stopped and all generated candidate regions are output.

[0041] Preferably, the total similarity of color, texture and size similarities between the regions calculated in the image acquisition and candidate region generation module is expressed as:

[0042] S(R i , R j )=ω1S color (R i , R j )+ω2S texture (R i , R j )+ω3S size (R i , R j )

[0043] In the formula, S(R i ,R j ) is the total similarity; R i , R j for each pair of adjacent regions. color (R i ,R j ) is a similarity measure based on color features; S texture 9R i ,R j ) is a similarity measure based on texture features; S size (R i ,R j ) is the similarity measure based on region size; ω1, ω2, ω3 are weight coefficients, the contribution of each feature to the total similarity.

[0044] The calculation formula for color similarity is:

[0045]

[0046] In the formula, For region R i The color mean of For region R j The color mean.

[0047] The calculation formula of texture similarity is as follows:

[0048]

[0049] In the formula, Region R i , R j The value of the kth bucket in the texture histogram.

[0050] The calculation formula for size similarity is:

[0051]

[0052] In the formula, |R i |、|R j |areas R i , R j area.

[0053] Preferably, the DERT model training and fine-tuning module uses the candidate region images generated by the selection search algorithm to perform self-supervised learning pre-training on the improved DERT model, and fine-tunes the DERT model through a small amount of labeled samples based on the self-supervised learning, specifically:

[0054] The candidate regions are divided into positive and negative samples according to their features, and each candidate region is standardized.

[0055] A self-supervised learning task including region reconstruction, target inference, and contrastive learning is designed and added to the DERT model.

[0056] The standardized candidate regions are input into the DERT model for training. The DERT model learns based on the self-supervised task and gradually adjusts the parameters so that the DERT model can capture the target information in the candidate regions.

[0057] The labeled samples are input into the DERT model that has been pre-trained through self-supervised learning, and a small batch learning strategy is adopted to fine-tune the DERT model with a small amount of labeled data.

[0058] On the other hand, the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the power transmission inspection small sample scene recognition method based on the selective search algorithm as described in any embodiment of the present invention when executing the computer program.

[0059] On the other hand, the present invention further provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for identifying a small sample scene for power transmission inspection based on a selective search algorithm as described in any embodiment of the present invention is implemented.

[0060] Compared with the prior art, the present invention has the following technical effects:

[0061] 1. This invention proposes an innovative training method that accelerates model convergence by cropping the original image. This method not only reduces the computational complexity but also increases the diversity of training samples by dividing a large-size image into multiple small blocks. By randomly combining and enhancing these candidate regions, the model is exposed to more diverse scenes and target distributions, thereby accelerating learning. This method not only significantly improves training efficiency, but also enhances the model's ability to perceive local features, which helps to improve the detection accuracy of small targets and targets in complex backgrounds.

[0062] 2. The present invention can significantly improve the efficiency of power transmission inspection by performing self-supervised learning pre-training and small sample fine-tuning training on the DERT model. In the self-supervised learning pre-training stage, the DERT model performs self-learning on a large number of unlabeled samples, extracts common visual features and structural information from them, and provides a strong basic capability for subsequent target detection tasks. Due to the high cost of labeling power transmission inspection data and the scarcity of data, self-supervised learning can enable the DERT model to obtain a basic understanding of various inspection scenarios in advance in the absence of labeled data. Through small sample fine-tuning training, the DERT model can be quickly optimized with only a small amount of labeled data, and its performance in power transmission inspection tasks can be fine-tuned, so that the DERT model can quickly adapt to and identify faults or anomalies of transmission lines in actual inspection scenarios. Overall, the accuracy and robustness of target detection in the inspection process are improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is an overall flow chart of the power transmission inspection small sample scene recognition method based on the selection search algorithm described in the present invention. DETAILED DESCRIPTION

[0064] In order to make the objectives, technical solutions and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in combination with specific embodiments of the present application and with reference to the accompanying drawings.

[0065] Embodiment 1

[0066] This embodiment provides a small sample scene recognition method for power transmission inspection based on a selection search algorithm. Figure 1 As shown, the following steps are included:

[0067] S1: Collect the original image of the transmission line inspection, generate candidate areas according to the original image by selecting the search algorithm, and extract the corresponding candidate areas from the original image as training samples according to the coordinate information of the candidate areas.

[0068] The selective search algorithm is a region recommendation algorithm for target detection, which is designed to be fast and has a high recall rate. The selective search algorithm described in this embodiment calculates the hierarchical grouping of similar regions based on the compatibility of color, texture, and size. As a preferred implementation of this embodiment, the step S1 of generating a candidate frame according to the original image by the selective search algorithm is specifically as follows:

[0069] S11: Perform superpixel segmentation on the original image and divide the original image into multiple regions {R1, R2…, R n}, where n is the number of superpixel regions. The pixels in each superpixel region are similar in color, texture, and shape.

[0070] S12: For each pair of adjacent regions, calculate the total similarity of color, texture, and size similarity between the regions.

[0071] As a preferred implementation of this embodiment, the total similarity of the color, texture, and size similarities between the calculated regions in step S12 is expressed as:

[0072] S(R i , R j )=ω1S color (R i , R j )+ω2S texture (R i , R j )+ω3S size (R i , R j )

[0073] In the formula, S(R i ,R j ) is the total similarity; R i , R j For each pair of adjacent regions; S color (R i ,R j ) is a similarity measure based on color features; S texture (Ri ,R j ) is a similarity measure based on texture features; S size (R i ,R j ) is the similarity measure based on region size; ω1, ω2, ω3 are weight coefficients, the contribution of each feature to the total similarity.

[0074] Color similarity calculates the mean value of the colors in the area and uses Euclidean distance to measure the color similarity. The calculation formula is as follows:

[0075]

[0076] In the formula, For region R i The color mean. For region R j The color mean.

[0077] Texture similarity measures the difference in texture patterns between two regions, usually using histogram comparison methods (such as histogram intersection, chi-square distance, etc.), and its calculation formula is as follows:

[0078]

[0079] In the formula, Region R i , R j The value of the kth bucket in the texture histogram.

[0080] Size similarity measures the size difference between two regions, and its calculation formula is:

[0081]

[0082] In the formula, |R i |、|R j |areas R i , R j area.

[0083] S13: If the total similarity of two adjacent regions exceeds a preset total similarity threshold, they are considered similar enough to be merged into a new region, thereby merging adjacent regions according to the similarity to generate a new candidate region.

[0084] S14: Repeat the merging process until the number of generated candidate regions reaches the preset target number, and then stop merging and output all generated candidate regions as input for the subsequent DERT model.

[0085] S2: Establish a DERT model, use the candidate regions generated by the selection search algorithm to perform self-supervised learning pre-training on the DERT model, and fine-tune the DERT model with a small number of labeled samples based on self-supervised learning to obtain a trained DERT model.

[0086] As a preferred implementation of this embodiment, in step S2, the improved DERT model is pre-trained by self-supervised learning using the candidate region images generated by the selection search algorithm, and the DERT model is fine-tuned by a small amount of labeled samples based on the self-supervised learning. Specifically,

[0087] S21: Divide the candidate region into positive and negative samples according to the characteristics of the candidate region, and perform standardization on each candidate region.

[0088] S22: Design self-supervised learning tasks including region reconstruction, target inference, and contrastive learning, and add them to the DERT model.

[0089] S23: The standardized candidate regions are input into the DERT model for training. The DERT model learns according to the self-supervised task and gradually adjusts the parameters so that the DERT model can capture the target information in the candidate region. Specifically, in the training phase of self-supervised learning, the DERT model can measure the difference between the candidate region reconstructed by the model and the original region through a reconstruction loss function (such as L2 loss or cosine similarity loss), predict the target category and position through target position and category loss functions (such as cross entropy loss and L1 loss), and minimize the feature difference between positive and negative samples through the contrast loss function. At the same time, the weight of each loss item is adjusted during the training process to balance the contribution of different tasks and prevent a certain task from excessively dominating the learning of the DERT model.

[0090] The contrast loss function can be expressed as:

[0091]

[0092] Where, L contrastive is the contrast loss function; is the mapping function, f(x) converts the input sample into a feature representation vector; x os 、x neg Represent positive and negative samples respectively; Δ is the preset distance threshold, which controls the maximum allowed distance difference between positive and negative samples.

[0093] In addition, to prevent background areas from dominating the training, an online hard example mining strategy can be used to select the most difficult to distinguish positive and negative samples for training.

[0094] S24: Input the labeled samples into the DERT model that has been pre-trained by self-supervised learning, and use a small batch learning strategy to fine-tune the DERT model using a small amount of labeled data. Specifically, in the small sample fine-tuning stage, the DERT model can perform category classification through cross entropy loss and bounding box regression through L1 loss + GIoU loss. At the same time, since the fine-tuning training data is small, more rounds of fine-tuning training can be performed.

[0095] After training is completed, the DERT model can be evaluated using the validation set or test set, and the AP (average precision) can be calculated to evaluate the detection performance and verify the performance of the DERT model on different data sets to ensure that the DERT model has good generalization ability. If the performance of the DERT model on a small sample data set is not ideal, you can try to increase the number of pre-training rounds or training samples in the fine-tuning stage.

[0096] S3: Collect the inspection images of the transmission line to be inspected in real time, and use the trained DERT model to perform transmission line inspection target detection.

[0097] Embodiment 2

[0098] Accordingly, this embodiment provides a power transmission inspection small sample scene recognition system based on a selective search algorithm. The system is used to implement the power transmission inspection small sample scene recognition method based on a selective search algorithm described in Example 1, including an image acquisition and candidate area generation module, a DERT model training and fine-tuning module, and a target detection module.

[0099] The image acquisition and candidate region generation module is used to acquire the original image of the transmission line inspection, generate the candidate region according to the original image by selecting the search algorithm, and extract the corresponding candidate region from the original image as the training sample according to the coordinate information of the candidate region.

[0100] The DERT model training and fine-tuning module is used to establish the DERT model. The candidate regions generated by the selection search algorithm are used to pre-train the DERT model through self-supervised learning. Based on the self-supervised learning, the DERT model is fine-tuned through a small number of labeled samples to obtain the trained DERT model.

[0101] The target detection module is used to collect the inspection images of the transmission line to be inspected in real time, and use the trained DERT model to perform transmission line inspection target detection.

[0102] As a preferred implementation of this embodiment, the image acquisition and candidate region generation module generates a candidate frame according to the original image through a selective search algorithm as follows:

[0103] Perform superpixel segmentation on the original image and divide the original image into multiple regions.

[0104] For each pair of adjacent regions, the total similarity of color, texture, and size similarity between the regions is calculated.

[0105] Adjacent regions are merged according to their similarity to generate new candidate regions.

[0106] The merging process is repeated until the number of generated candidate regions reaches the preset target number, then the merging is stopped and all generated candidate regions are output.

[0107] As a preferred implementation of this embodiment, the total similarity of color, texture, and size similarities between the calculated regions in the image acquisition and candidate region generation module is expressed as:

[0108] S(R i , R j )=ω1S color (R i , R j )+ω2S texture (R i , R j )+ω3S size (R i , R j )

[0109] In the formula, S(R i ,R j ) is the total similarity; R i , R j For each pair of adjacent regions; S color (R i ,R j ) is a similarity measure based on color features; S texture (R i ,R j ) is a similarity measure based on texture features; S size (R i ,R j ) is the similarity measure based on region size; ω1, ω2, ω3 are weight coefficients, the contribution of each feature to the total similarity.

[0110] The calculation formula for color similarity is:

[0111]

[0112] In the formula, For region R i The color mean of For region R j The color mean.

[0113] The calculation formula of texture similarity is as follows:

[0114]

[0115] In the formula, Region R i , R j The value of the kth bucket in the texture histogram.

[0116] The calculation formula for size similarity is:

[0117]

[0118] In the formula, |R i |、|R j |areas R i , R j area.

[0119] As a preferred implementation of this embodiment, the DERT model training and fine-tuning module uses the candidate region images generated by the selection search algorithm to perform self-supervised learning pre-training on the improved DERT model, and fine-tunes the DERT model through a small amount of labeled samples based on self-supervised learning. Specifically:

[0120] The candidate regions are divided into positive and negative samples according to their features, and each candidate region is standardized.

[0121] A self-supervised learning task including region reconstruction, target inference, and contrastive learning is designed and added to the DERT model.

[0122] The standardized candidate regions are input into the DERT model for training. The DERT model learns based on the self-supervised task and gradually adjusts the parameters so that the DERT model can capture the target information in the candidate regions.

[0123] The labeled samples are input into the DERT model that has been pre-trained through self-supervised learning, and a small batch learning strategy is adopted to fine-tune the DERT model with a small amount of labeled data.

[0124] Embodiment 3

[0125] This embodiment provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the power transmission inspection small sample scene recognition method based on a selective search algorithm as described in any embodiment of the present invention when executing the computer program.

[0126] Embodiment 4

[0127] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method for identifying a small sample scenario for power transmission inspection based on a selective search algorithm as described in any embodiment of the present invention is implemented.

[0128] 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.

[0129] 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.

[0130] 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.

[0131] 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.

[0132] 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 small sample scene recognition method for power transmission inspection based on a selective search algorithm, characterized in that: The following steps are involved: S1: Collect the original image of the transmission line inspection, generate candidate areas according to the original image by selecting the search algorithm, and extract the corresponding candidate areas from the original image as training samples according to the coordinate information of the candidate areas; S2: Establish a DERT model, use the candidate regions generated by the selection search algorithm to perform self-supervised learning pre-training on the DERT model, and fine-tune the DERT model with a small number of labeled samples based on self-supervised learning to obtain a trained DERT model; S3: Collect the inspection images of the transmission line to be inspected in real time, and use the trained DERT model to perform transmission line inspection target detection.

2. The method for identifying small sample scenarios for power transmission inspection based on a selective search algorithm according to claim 1 is characterized in that: In step S1, generating a candidate frame according to the original image by using a search algorithm is specifically as follows: S11: performing superpixel segmentation on the original image to divide the original image into multiple regions; S12: For each pair of adjacent regions, calculate the total similarity of color, texture, and size similarity between the regions; S13: Merge adjacent regions according to similarity to generate new candidate regions; S14: Repeat the merging process until the number of generated candidate regions reaches a preset target number, then stop merging and output all generated candidate regions.

3. The power transmission inspection small sample scene recognition method based on the selective search algorithm according to claim 2 is characterized in that: The total similarity of the color, texture and size similarities between the calculated regions in step S12 is expressed as: S(R i ,R j )=ω1S color (R i ,R j )+ω2S texture (R i ,R j )+ω3S size (R i ,R j ) In the formula, S(R i ,R j ) is the total similarity; R i , R j For each pair of adjacent regions; S color (R i ,R j ) is a similarity measure based on color features; S texture (R i ,R j ) is a similarity measure based on texture features; S size (R i ,R j ) is the similarity measure based on region size; ω1, ω2, ω3 are weight coefficients, the contribution of each feature to the total similarity; The calculation formula for color similarity is: In the formula, For region R i The color mean of For region R j The color mean of The calculation formula of texture similarity is as follows: In the formula, Region R i , R j The value of the kth bucket in the texture histogram; The calculation formula for size similarity is: In the formula, |R i |、|R j |areas R i , R j area.

4. The power transmission inspection small sample scene recognition method based on the selective search algorithm according to claim 1 is characterized in that: In step S2, the improved DERT model is pre-trained by self-supervised learning using the candidate region images generated by the selection search algorithm. Based on the self-supervised learning, the DERT model is fine-tuned by a small amount of labeled samples. Specifically, the following steps are performed: S21: Divide the candidate regions into positive and negative samples according to their features, and perform standardization on each candidate region; S22: Design self-supervised learning tasks including region reconstruction, target inference, and contrastive learning, and add them to the DERT model; S23: The standardized candidate regions are input into the DERT model for training. The DERT model learns according to the self-supervision task and gradually adjusts the parameters so that the DERT model can capture the target information in the candidate regions. S24: Input the labeled samples into the DERT model that has been pre-trained through self-supervised learning, adopt a small batch learning strategy, and use a small amount of labeled data to fine-tune the DERT model.

5. A small sample scene recognition system for power transmission inspection based on a selective search algorithm, characterized in that: It includes image acquisition and candidate region generation module, DERT model training and fine-tuning module and target detection module; The image acquisition and candidate region generation module is used to acquire the original image of the transmission line inspection, generate the candidate region according to the original image by selecting the search algorithm, and extract the corresponding candidate region from the original image as the training sample according to the coordinate information of the candidate region; The DERT model training and fine-tuning module is used to establish the DERT model. The candidate regions generated by the selection search algorithm are used to pre-train the DERT model through self-supervised learning. Based on the self-supervised learning, the DERT model is fine-tuned through a small number of labeled samples to obtain the trained DERT model. The target detection module is used to collect the inspection images of the transmission line to be inspected in real time, and use the trained DERT model to perform transmission line inspection target detection.

6. The power transmission inspection small sample scene recognition system based on the selective search algorithm according to claim 5 is characterized in that: The image acquisition and candidate region generation module generates a candidate frame according to the original image through a selective search algorithm as follows: Perform superpixel segmentation on the original image to divide the original image into multiple regions; For each pair of adjacent regions, the total similarity of color, texture, and size similarity between the regions is calculated; Merge adjacent regions according to similarity to generate new candidate regions; The merging process is repeated until the number of generated candidate regions reaches the preset target number, then the merging is stopped and all generated candidate regions are output.

7. The power transmission inspection small sample scene recognition system based on the selective search algorithm according to claim 6 is characterized in that: The total similarity of the color, texture, and size similarities between the regions calculated in the image acquisition and candidate region generation module is expressed as: S(R i ,R j )=ω1S color (R i ,R j )+ω2S texture (R i ,R j )+ω3S size (R i ,R j ) In the formula, S(R i ,R j ) is the total similarity; R i , R j For each pair of adjacent regions; S color (R i ,R j ) is a similarity measure based on color features; S texture 9R i ,R j ) is a similarity measure based on texture features; S size (R i ,R j ) is the similarity measure based on region size; ω1, ω2, ω3 are weight coefficients, the contribution of each feature to the total similarity; The calculation formula for color similarity is: In the formula, For region R i The color mean of For region R j The color mean of The calculation formula of texture similarity is as follows: In the formula, Region R i , R j The value of the kth bucket in the texture histogram; The calculation formula for size similarity is: In the formula, |R i |、|R j |areas R i , R j area.

8. The power transmission inspection small sample scene recognition system based on the selective search algorithm according to claim 5 is characterized in that: In the DERT model training and fine-tuning module, the candidate region images generated by the selection search algorithm are used to perform self-supervised learning pre-training on the improved DERT model. Based on the self-supervised learning, the DERT model is fine-tuned by a small number of labeled samples. Specifically, The candidate regions are divided into positive and negative samples according to their characteristics, and each candidate region is standardized; Design self-supervised learning tasks including region reconstruction, target inference, and contrastive learning, and add them to the DERT model; The standardized candidate regions are input into the DERT model for training. The DERT model learns based on the self-supervision task and gradually adjusts the parameters so that the DERT model can capture the target information in the candidate regions. The labeled samples are input into the DERT model that has been pre-trained through self-supervised learning, and a small batch learning strategy is adopted to fine-tune the DERT model with a small amount of labeled data.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the power transmission inspection small sample scene recognition method based on a selective search algorithm as described in any one of claims 1 to 4 when executing the computer program.

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 identifying small sample scenarios for power transmission inspection based on a selective search algorithm as claimed in any one of claims 1 to 4 is implemented.

Images