An intelligent detection method and system for circuit defects based on a drone platform
By combining a lightweight deep convolutional neural network and a Transformer model, a circuit defect intelligent detector has been developed, which solves the problems of insufficient efficiency and accuracy in circuit defect detection on UAV platforms and achieves efficient and accurate circuit defect detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2023-02-08
- Publication Date
- 2026-06-26
AI Technical Summary
Existing intelligent circuit defect detection technologies based on UAV platforms suffer from insufficient detection efficiency and accuracy. In particular, the local receptive field of deep convolutional neural networks limits further improvement in detection accuracy.
A circuit defect intelligent detector is constructed using a lightweight deep convolutional neural network and a Transformer model. The lightweight deep convolutional neural network improves the model's detection efficiency, while the global receptive field characteristics of the Transformer model enhance the detection accuracy.
This technology enables rapid and accurate detection of circuit defects on a drone platform, improving detection efficiency and accuracy and solving the problems of insufficient efficiency and accuracy in existing technologies.
Smart Images
Figure CN115984245B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent circuit defect detection technology, and in particular to an intelligent circuit defect detection method and system based on an unmanned aerial vehicle (UAV) platform. Background Technology
[0002] In recent years, with rapid economic development, power transmission networks have become increasingly sophisticated, bringing more and more complex tasks to circuit inspection. The ever-increasing workload cannot be completed manually. Intelligent circuit defect detection aims to identify potential circuit defects in each frame of the inspection image, offering significantly higher efficiency compared to manual inspection. Therefore, intelligent circuit defect detection has become the best solution for improving circuit inspection efficiency and ensuring reliable power supply for civilian and industrial use. Unmanned aerial vehicle (UAV) platforms possess unparalleled flexibility compared to other carriers. Intelligent circuit defect detection based on UAV platforms can perform defect detection on circuits in various complex environments (high altitude, strong winds, high voltage) while ensuring the safety of inspection personnel. Therefore, research on intelligent circuit defect detection technology based on UAV platforms has extremely high application value and research significance.
[0003] Currently, research on intelligent circuit defect detection based on UAV platforms is relatively scarce. However, apart from the limitation on model size imposed by the UAV platform, the other requirements for intelligent circuit defect detection based on UAV platforms are the same as those for intelligent circuit defect detection. Therefore, the two tasks are highly similar and can be discussed from the perspective of intelligent circuit defect detection. To achieve high detection accuracy, most existing intelligent circuit defect detection systems use deep convolutional neural networks (DCNNs) to design models. These DCNNs extract image features, and then the target detector designed based on DCNNs is trained using labeled data. Deep learning combined with extensive data training aims to achieve high detection accuracy. However, the local receptive field of DCNNs limits further improvements in the detection accuracy of DCNN-based models.
[0004] Therefore, there is an urgent need for an intelligent circuit defect detection technology based on an unmanned aerial vehicle (UAV) platform that offers high detection efficiency and accuracy. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for intelligent detection of circuit defects based on an unmanned aerial vehicle (UAV) platform, which can improve both detection efficiency and accuracy compared to existing intelligent circuit defect detection methods.
[0006] To achieve the above objectives, the present invention provides the following solution:
[0007] A method for intelligent detection of circuit defects based on an unmanned aerial vehicle (UAV) platform, the method comprising:
[0008] Obtain inspection images of circuits using a drone platform;
[0009] Using the inspection image as input, a circuit defect intelligent detector is used to determine the category and location of each defect in the inspection image; the circuit defect intelligent detector is composed of a lightweight deep convolutional neural network and a Transformer model.
[0010] In some embodiments, the intelligent circuit defect detector includes a feature extraction module and a target detector connected in sequence; the feature extraction module is a lightweight deep convolutional neural network; and the target detector is a Transformer model.
[0011] In some embodiments, the lightweight deep convolutional neural network includes a plurality of convolutional layers connected in sequence.
[0012] In some embodiments, before determining the category and location of each defect in the inspection image using the circuit defect intelligent detector as input, the circuit defect intelligent detection method further includes: training an initial model using a dataset to obtain a circuit defect intelligent detector; the dataset includes multiple sample images and a label for each sample image, the label including the true category and true location of each defect in the sample image; the initial model and the circuit defect intelligent detector have the same model structure.
[0013] In some embodiments, the loss function used when training the initial model using the dataset is:
[0014] L = L cls +L loc ;
[0015] Where L is the total loss; L cls For classification loss; L loc To pinpoint the loss.
[0016] In some embodiments, the formula for calculating the classification loss is:
[0017]
[0018] Where N is the number of predicted bounding boxes based on anchor prediction; Pos is the set of positive predicted bounding boxes; Let be the classification score of the defect in the p-th positive predicted bounding box belonging to class a, where 'a' is the true class of the true bounding box that matches the p-th positive predicted bounding box; Neg is the set of negative predicted bounding boxes. The classification score for the defect in the nth negative prediction bounding box belonging to class 0.
[0019] In some embodiments, the formula for calculating the positioning loss is:
[0020]
[0021] Where N is the number of predicted bounding boxes based on anchor prediction; M is the number of ground truth bounding boxes; d ij A binary variable representing whether the loss needs to be calculated between the predicted coordinates of the i-th predicted bounding box and the ground truth coordinates of the j-th ground truth bounding box; smoothL1 is the loss function; l i Let gt be the predicted coordinates of the i-th predicted bounding box; j Let be the true coordinates of the j-th true bounding box.
[0022] In some embodiments, after determining the category and location of each defect in the inspection image using a circuit defect intelligent detector, the circuit defect intelligent detection method further includes removing defects whose category classification score is lower than a preset threshold.
[0023] In some embodiments, the feature extraction module is a pre-trained lightweight deep convolutional neural network.
[0024] A circuit defect intelligent detection system based on an unmanned aerial vehicle (UAV) platform, the circuit defect intelligent detection system comprising:
[0025] The image acquisition module is used to acquire inspection images obtained by using a drone platform to inspect circuits.
[0026] The detection module is used to determine the category and location of each defect in the inspection image by using the intelligent circuit defect detector as input; the intelligent circuit defect detector is composed of a lightweight deep convolutional neural network and a Transformer model.
[0027] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0028] This invention provides a method and system for intelligent circuit defect detection based on an unmanned aerial vehicle (UAV) platform. It designs an intelligent circuit defect detector composed of a lightweight deep convolutional neural network (DCNN) and a Transformer model. By employing a lightweight DCNN, the model's detection efficiency is improved. Furthermore, the Transformer model, with its global receptive field compared to DCNNs, effectively overcomes the limitation on detection accuracy imposed by the local receptive field of DCNNs, thus enhancing detection accuracy. Based on this intelligent circuit defect detector, the inspection images obtained from circuit inspections using an UAV platform can be quickly and accurately determined to identify the type and location of each defect in the inspection images. Compared to existing intelligent circuit defect detection methods, this invention simultaneously improves both detection efficiency and accuracy. Attached Figure Description
[0029] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0030] Figure 1 This is a flowchart of the intelligent circuit defect detection method provided in Embodiment 1 of the present invention;
[0031] Figure 2 This is a schematic diagram of the model structure of the intelligent circuit defect detector provided in Embodiment 1 of the present invention;
[0032] Figure 3 This is a schematic diagram of the pre-training process provided in Embodiment 1 of the present invention;
[0033] Figure 4 This is a schematic diagram of the initial model training process provided in Embodiment 1 of the present invention;
[0034] Figure 5 This is a system block diagram of the intelligent circuit defect detection system provided in Embodiment 2 of the present invention. Detailed Implementation
[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0036] The purpose of this invention is to provide a method and system for intelligent detection of circuit defects based on an unmanned aerial vehicle (UAV) platform, which can improve both detection efficiency and accuracy compared to existing intelligent circuit defect detection methods.
[0037] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0038] Example 1:
[0039] This embodiment provides a method for intelligent detection of circuit defects based on an unmanned aerial vehicle (UAV) platform, such as... Figure 1 As shown, the intelligent circuit defect detection method includes:
[0040] S1: Obtain inspection images of circuits using a drone platform;
[0041] This embodiment utilizes a drone platform to inspect circuits and takes pictures of the circuits during the inspection process to obtain inspection images. Using a drone platform to inspect circuits is a highly efficient method that also ensures the safety of inspection personnel.
[0042] S2: Using the inspection image as input, the type and location of each defect in the inspection image are determined by a circuit defect intelligent detector; the circuit defect intelligent detector is composed of a lightweight deep convolutional neural network and a Transformer model.
[0043] This embodiment takes each frame of inspection image as input and uses a circuit defect intelligent detector to detect each defect in each frame of inspection image, and determines the category and location of each defect. The category can include open circuit, short circuit, short circuit, grounding, and wiring error, etc. The location can be described by the coordinates of the four vertices of the bounding box surrounding the defect: top, bottom, left, and right.
[0044] like Figure 2 As shown, the circuit defect intelligent detector in this embodiment includes a feature extraction module and a target detector connected in sequence. The feature extraction module is a lightweight deep convolutional neural network, which may include multiple convolutional layers connected in sequence. Figure 2 In this model, the lightweight deep convolutional neural network includes four convolutional layers. However, those skilled in the art will understand that the number of convolutional layers can be designed according to actual needs in practical applications. The target detector is a Transformer model. During detection, the inspection image is input into the feature extraction module, which then sends the extracted image features to the target detector for circuit defect detection, determining the category and location of each defect in the inspection image.
[0045] Preferably, the feature extraction module in this embodiment can also be a pre-trained lightweight deep convolutional neural network. Pre-training means training the model parameters of the lightweight deep convolutional neural network in advance using large-scale general data, such as... Figure 3 As shown, the pre-training process involves training the model parameters of a lightweight deep convolutional neural network using videos with labeled information. The labeled information can include the true class and the true location. Existing mature training procedures can be used, which will not be elaborated upon here. Pre-training can accelerate the training speed of lightweight deep convolutional neural networks and improve their performance.
[0046] Considering the limitations imposed by the UAV platform on model size, this embodiment employs a lightweight deep convolutional neural network as the feature extraction module when designing the intelligent circuit defect detector for intelligent circuit defect detection based on the UAV platform. This significantly improves the model's detection efficiency. Furthermore, the local receptive field of deep convolutional neural networks limits further improvements in their detection accuracy. In contrast, the Transformer model possesses a global receptive field, effectively addressing the limitations imposed by the local receptive field of deep convolutional neural networks on detection accuracy. Therefore, this embodiment uses the Transformer model as the target detector in the design of the intelligent circuit defect detector for intelligent circuit defect detection based on the UAV platform, which greatly enhances the model's detection accuracy. Therefore, the advantages of the circuit defect intelligent detector in this embodiment are: (1) A lightweight circuit defect intelligent detector is constructed using a lightweight deep convolutional neural network, which improves the model detection efficiency and is conducive to deployment on UAV platforms; (2) To address the problem of limited detection accuracy of the circuit defect detection model using a pure deep convolutional neural network, a model structure design for a circuit defect intelligent detector combining a lightweight deep convolutional neural network and a Transformer model is proposed, which improves the model detection accuracy through the global receptive field mechanism of the Transformer model. Therefore, the circuit defect intelligent detector designed in this embodiment can simultaneously take into account both detection efficiency and detection accuracy. Compared with existing circuit defect intelligent detection, the method of UAV-based circuit defect intelligent detection using the above-mentioned circuit defect intelligent detector in this embodiment can simultaneously improve detection efficiency and detection accuracy.
[0047] To address the current lack of intelligent circuit defect detection methods based on UAV platforms, this embodiment provides an end-to-end intelligent circuit defect detection method based on UAV platforms, employing lightweight deep convolutional neural networks and Transformer models. By utilizing lightweight deep convolutional neural networks and Transformer models as intelligent circuit defect detectors, this method can improve both model detection efficiency and accuracy.
[0048] Preferably, after S2, the intelligent circuit defect detection method of this embodiment further includes: removing defects whose classification scores are lower than a preset threshold, thereby enabling more accurate identification of circuit defects in the inspection images.
[0049] Prior to S2, the intelligent circuit defect detection method in this embodiment also includes a training process for training an initial model using a dataset to obtain an intelligent circuit defect detector, such as... Figure 4 As shown, the training process may include:
[0050] (1) Design the circuit defect detection model structure:
[0051] This embodiment designs an end-to-end lightweight deep convolutional neural network and Transformer model network structure to meet the task requirements of intelligent circuit defect detection on UAV platforms. The initial model is identical to the model structure of the intelligent circuit defect detector, which includes a feature extraction module and a target detector connected in sequence. The feature extraction module is a lightweight deep convolutional neural network, and the target detector is a Transformer model. The detection accuracy is improved by the global receptive field mechanism of the Transformer model.
[0052] (2) Model training:
[0053] A dataset is constructed, which includes multiple sample images and labels for each sample image. The labels include the true category and true location of each defect in the sample image. The network model structure designed in (1) (i.e., the initial model) is trained and optimized using the sample images with labels in the dataset to obtain the trained model parameters. The initial model using the trained model parameters is the circuit defect intelligent detector that balances detection efficiency and detection accuracy.
[0054] Specifically, during the training process, sample images are... (Where H is the image height, W is the image width, and 3 is the number of image channels; in this embodiment, the sample image can be an RGB three-channel image.) The image is input to the feature extraction module, which extracts rich image features. These extracted features are then input to the object detector, which includes an encoder and a decoder. The object detector uses the encoder to perform self-attention calculations on the image features, enhancing them. The enhanced image features are then input to the decoder, which predicts the classification score set and the location coordinate set. (Classification score set) Where N is the number of predicted bounding boxes based on anchor prediction in the sample image, and C is the total number of categories in the dataset. Let L be the classification score of the defect in the i-th predicted bounding box belonging to class c, where i = 1, 2, ..., N, c = 0, 2, ..., C, and c = 0 refers to any class other than those in the dataset. The set of location coordinates L = {l i}∈R N×2 , where l i Let be the predicted coordinates of the i-th predicted bounding box. The predicted coordinates may include the coordinates of the lower left vertex and the upper right vertex of the predicted bounding box. It should be noted that the sum of the classification scores of the defects in the i-th predicted bounding box belonging to each class is 1.
[0055] Then, the initial model is learned and optimized using the label information. The labels include the true category and true location of each of the M true bounding boxes. If the true category of the defect in the j-th true bounding box is class c, then the classification score of the defect in the j-th true bounding box belonging to class c is 1, and the classification score of the defect in the j-th true bounding box belonging to any category other than class c is 0. The true location includes the true coordinates of the j-th true bounding box, which may include the coordinates of the lower left vertex and the upper right vertex of the true bounding box. In this embodiment, the predicted classification score set, the location coordinate set, and the labels of the sample images are used as input. The initial model is learned and optimized using the loss function to obtain intermediate model parameters. This completes one iteration. The iteration is continued until the total loss calculated using the loss function no longer changes significantly. Then the iteration ends. The intermediate model parameters at this point are the trained model parameters, and the intelligent circuit defect detector can be obtained.
[0056] The loss function used is:
[0057] L = L cls +L loc ;
[0058] Where L is the total loss; L cls For classification loss; L loc To pinpoint the loss.
[0059] The formula for calculating classification loss is:
[0060]
[0061] Where N is the number of predicted bounding boxes based on anchor prediction; Pos is the set of positive predicted bounding boxes; Let be the classification score of the defect in the p-th positive predicted bounding box belonging to class a, where 'a' is the true class of the true bounding box that matches the p-th positive predicted bounding box; Neg is the set of negative predicted bounding boxes. The classification score for the defect in the nth negative prediction bounding box belonging to class 0.
[0062] In this embodiment, predicted bounding boxes are divided into positive predicted bounding boxes and negative predicted bounding boxes. All positive predicted bounding boxes form a positive predicted bounding box set, and all negative predicted bounding boxes form a negative predicted bounding box set. The sum of the positive predicted bounding boxes in the positive predicted bounding box set and the negative predicted bounding boxes in the negative predicted bounding box set is N. The partitioning method is as follows: For each ground truth bounding box, the intersection-union ratio (IUR) of each predicted bounding box with the ground truth bounding box is calculated. Predicted bounding boxes with IUR greater than a preset threshold are selected as candidate bounding boxes. The candidate bounding box with the maximum IUR is selected as the matching bounding box that matches the ground truth bounding box. The matching bounding boxes that match the ground truth bounding boxes can be determined in the above manner. If the IUR of all predicted bounding boxes is less than or equal to the preset threshold, then the ground truth bounding box has no matching bounding box. The matching bounding boxes that can match the ground truth bounding boxes belong to the positive predicted bounding boxes, and the remaining predicted bounding boxes belong to the negative predicted bounding boxes, thereby constructing the positive predicted bounding box set and the negative predicted bounding box set.
[0063] The formula for calculating positioning loss is:
[0064]
[0065] Where, mean represents the average value, N is the number of predicted bounding boxes based on anchor prediction, M is the number of true bounding boxes, and d ij A binary variable representing whether the loss needs to be calculated between the predicted coordinates of the i-th predicted bounding box and the ground truth coordinates of the j-th ground truth bounding box; smoothL1 is the loss function; l i Let gt be the predicted coordinates of the i-th predicted bounding box; j Let be the true coordinates of the j-th true bounding box.
[0066] d ijThe determination method is as follows: For the j-th true bounding box, calculate the intersection-union ratio (IU) between each predicted bounding box and the true bounding box. Select the predicted bounding boxes with IU ranking in the top Q positions as the decision bounding boxes, in descending order. In this embodiment, it is necessary to calculate the loss between the predicted coordinates of the decision bounding box and the true coordinates of the j-th true bounding box. Assuming the decision bounding boxes are the 1st, 2nd, and 5th predicted bounding boxes, then in the actual calculation, d 1j d 2j d 5j One is 1, and the rest are 0.
[0067] It should be noted that the above loss function is for a single sample image. During training, the total loss for each sample image is calculated using the above loss function, and then the sum of the total losses for each sample image is calculated to obtain the total loss for this iteration.
[0068] (3) Model testing:
[0069] Circuit images captured by a drone platform are input into a circuit defect intelligent detector as test images. The detector detects the category and location of each defect in the test image. The detection results in the test image can then be post-processed. Defects with classification scores below a preset threshold are removed using NMS (non-maximum suppression). This process yields the category and location of each defect in the final test image, thus determining the category and location of the circuit defects contained in the test image.
[0070] Specifically, during testing, the test image is input into the feature extraction module, which extracts image features. These extracted features are then input into the encoder of the target detector. The self-attention calculation of the convolutional neural network in the encoder further enhances the image features. The enhanced features are then input into the decoder. The convolutional neural network in the decoder predicts and generates a set of classification scores and location coordinates for N predicted bounding boxes. Each of the N predicted bounding boxes corresponds to a defect. The category corresponding to the maximum classification score of the i-th predicted bounding box is selected as the category of the i-th defect. The location of the i-th defect (including four coordinates: top, bottom, left, and right) is determined based on the predicted coordinates of the i-th predicted bounding box. In this embodiment, the bounding box is rectangular. This yields a preliminary detection result of the defect location and category for the entire test image. Defects with classification scores below a preset threshold in the preliminary detection result are removed, resulting in the final category and location of each circuit defect.
[0071] After the model testing is completed, the intelligent circuit defect detector is directly ported to the UAV platform. By leveraging the flexibility of the UAV platform, it can be deployed flexibly, thereby realizing intelligent circuit defect detection based on the UAV platform.
[0072] Example 2:
[0073] This embodiment provides an intelligent circuit defect detection system based on an unmanned aerial vehicle (UAV) platform, such as... Figure 5 As shown, the intelligent circuit defect detection system includes:
[0074] Image acquisition module M1 is used to acquire inspection images obtained by using a drone platform to inspect circuits.
[0075] The detection module M2 is used to determine the category and location of each defect in the inspection image by using the circuit defect intelligent detector as input; the circuit defect intelligent detector is composed of a lightweight deep convolutional neural network and a Transformer model.
[0076] This embodiment provides an intelligent circuit defect detection system based on an unmanned aerial vehicle (UAV) platform. Addressing the specific needs of circuit defect detection on an UAV platform, it employs a model design combining a lightweight deep convolutional neural network and a Transformer model, balancing detection efficiency and accuracy. Compared to existing intelligent circuit defect detection systems, it simultaneously improves both detection efficiency and accuracy.
[0077] Each embodiment in this specification focuses on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be found in the method section.
[0078] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for intelligent detection of circuit defects based on an unmanned aerial vehicle (UAV) platform, characterized in that, The intelligent circuit defect detection method includes: Obtain inspection images of circuits using a drone platform; Using the inspection image as input, a circuit defect intelligent detector is used to determine the category and location of each defect in the inspection image. The circuit defect intelligent detector is composed of a lightweight deep convolutional neural network and a Transformer model. The circuit defect intelligent detector includes a feature extraction module and a target detector connected in sequence. The feature extraction module is a lightweight deep convolutional neural network, which includes multiple convolutional layers connected in sequence. The target detector is a Transformer model. The initial model is trained using a dataset to obtain a circuit defect intelligent detector; the dataset includes multiple sample images and a label for each sample image, the label including the true category and true location of each defect in the sample image; the initial model and the circuit defect intelligent detector have the same model structure; The loss function used when training the initial model using the dataset is: ； in, Total loss; For classification loss; To locate the loss; The formula for calculating the classification loss is: ； in, This represents the number of predicted bounding boxes based on anchor prediction. This is the set of positive predicted bounding boxes; For the first The defect in the positive prediction bounding box belongs to the first... Class classification score, In order to be with the first The true category of the true bounding box that matches the positive predicted bounding box; This is the set of negative predicted bounding boxes; For the first The defect in each negative predicted bounding box belongs to the classification score of class 0, where 0 represents any class other than those in the dataset. For each ground truth bounding box, calculate the intersection-union ratio (IUR) between each predicted bounding box and the ground truth bounding box. Select the predicted bounding boxes with IUR greater than a preset threshold as candidate bounding boxes. Select the candidate bounding box corresponding to the maximum IUR as the matching bounding box that matches the ground truth bounding box. The matching bounding boxes that match the ground truth bounding boxes are positive predicted bounding boxes, and the remaining predicted bounding boxes are negative predicted bounding boxes. The formula for calculating the positioning loss is: ； in, This represents the number of predicted bounding boxes based on anchor prediction. The number of true bounding boxes; To indicate whether the first step needs to be calculated The predicted coordinates of the predicted bounding box and the predicted coordinates of the predicted bounding box The loss is a binary variable representing the true coordinates of a real bounding box; The loss function; For the first The predicted coordinates of each predicted bounding box; For the first The true coordinates of a real bounding box; For the For each of the 10 true bounding boxes, calculate the intersection-union ratio (IU) between the predicted bounding box and the true bounding box. Select the predicted bounding boxes with the highest IU in descending order as the decision bounding boxes. Then, calculate the predicted coordinates of the decision bounding boxes and the 10th true bounding box. The loss is the true coordinates of the true bounding box.
2. The intelligent circuit defect detection method according to claim 1, characterized in that, After using the inspection image as input and determining the category and location of each defect in the inspection image using a circuit defect intelligent detector, the circuit defect intelligent detection method further includes: removing defects whose category classification score is lower than a preset threshold.
3. The intelligent circuit defect detection method according to claim 1, characterized in that, The feature extraction module is a pre-trained lightweight deep convolutional neural network.
4. A circuit defect intelligent detection system based on an unmanned aerial vehicle (UAV) platform, employing the circuit defect intelligent detection method based on an UAV platform as described in any one of claims 1-3, characterized in that, The intelligent circuit defect detection system includes: The image acquisition module is used to acquire inspection images obtained by using a drone platform to inspect circuits. The detection module is used to determine the category and location of each defect in the inspection image by using the intelligent circuit defect detector as input; the intelligent circuit defect detector is composed of a lightweight deep convolutional neural network and a Transformer model.