Defect identification method and device based on image processing and intelligent substation patrol system

Abstract

The invention relates to a defect identification method and device based on image processing and an intelligent substation patrol system, and belongs to the technical field of image processing. The method comprises the following steps: inputting an image to be detected into a positive sample detection model, judging whether the image to be detected has defects or not, inputting the image to be detected with the defects into a negative sample detection model, and obtaining the defect type of the image to be detected; according to the positive sample detection model, a positive sample with the maximum similarity with the to-be-detected image is selected from a positive sample database to serve as a reference image, the reference image and the to-be-detected image are compared and analyzed, and a comparison result is output; the negative sample detection model is obtained by training a negative sample set composed of image samples of different types of defects and corresponding defect types. According to the method, on the basis of reducing the identification complexity, the missing report rate and the false report rate are reduced, the reliability of a detection result is improved, the number of manually audited pictures is reduced, and the operation, maintenance and management cost of the system is effectively reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a defect recognition method, device and substation intelligent inspection system based on image processing. Background Art

[0002] With the continuous expansion of the scale and the increasing complexity of the power system, as an important part of the power system, the safe, stable and efficient operation of the substation is of crucial importance. Traditional substation inspection methods usually rely on manual inspection and manual detection, which have problems such as large workload, low efficiency and high safety risks. In recent years, with the rapid development of information technology and the rise of artificial intelligence, intelligent inspection technology has gradually been applied to the operation and maintenance management of substations. The substation remote intelligent inspection system has gradually developed from the pilot construction stage to the small-scale commercial stage, and has initially realized functions such as meter-free transcription, automatic data analysis and comparison, and automatic identification of appearance abnormalities, which has played an important auxiliary role in improving the inspection quality of operation and maintenance personnel. However, from the actual application effect of the substation remote intelligent inspection system, there are still some problems.

[0003] Due to the lack of missing sample data, traditional image recognition methods for negative samples (image samples containing defects) have many cases of missed reports and false alarms. Operation and maintenance personnel need to conduct secondary verification on all inspection pictures, resulting in a large workload and affecting the enthusiasm of operation and maintenance personnel. To solve the defects existing in the negative sample detection method, a Chinese patent application with publication number CN117576077A discloses a defect detection method. When training the detection model, positive samples (image samples without defects) are added to the image samples, and the positive samples and negative samples are used together as training samples to train the detection model. Then, the image to be detected is input into the trained detection model to obtain the final detection result. However, the detection model used in this detection method is difficult to train, resulting in a complex defect recognition process, and there are still cases of missed reports and false alarms. Summary of the Invention

[0004] The purpose of the present invention is to provide a defect recognition method, device and substation intelligent inspection system based on image processing to solve the problems of missed reports and false alarms existing in the existing defect detection methods and the unreliable detection results.

[0005] The defect recognition method based on image processing provided by the present invention to solve the above technical problems is as follows: input the acquired image to be detected into the positive sample detection model to determine whether there are defects in the image to be detected, and input the image to be detected with defects into the negative sample detection model to obtain the defect type of the image to be detected; the positive sample detection model is: select the positive sample with the highest similarity to the image to be detected from the constructed positive sample database as the reference image, compare and analyze the reference image and the image to be detected, and output the comparison result; the negative sample detection model is trained through a negative sample set, and the negative sample set is composed of image samples containing different types of defects and the corresponding defect types.

[0006] Further, the method for constructing the positive sample database is: collect image samples under different lighting conditions and shooting angles in the application scenario, perform image cleaning on the image samples, and remove the image samples with signal loss, or the image samples whose image clarity, brightness, or preset position deviation do not meet the image recognition conditions, and establish a positive sample database with the image acquisition device point as the unit.

[0007] Further, when the image sample has at least one of the following situations, it is considered that there is signal loss: 1) the image is blurred due to incorrect focal length of the shooting device or movement of the shooting device; 2) noise is introduced during the acquisition, transmission, or processing process, and the noise is greater than the noise threshold; 3) the image is compressed and distorted; 4) there is a color screen or black screen phenomenon.

[0008] Further, the detection of image clarity includes four parts: abnormal brightness detection, image noise level estimation, abnormal image color cast detection, and image blurriness estimation.

[0009] Further, the negative sample detection model includes a feature extraction network, a candidate region extraction and fusion algorithm, and a classification network.

[0010] Further, the candidate region extraction and fusion algorithm includes four algorithms: MCG, Selective Search, Bing, and EdgeBox. Generate target candidate regions through the four algorithms respectively, merge the target candidate regions through non-maximum suppression, and retain the candidate region with the highest confidence.

[0011] Further, the steps of the image noise level estimation include: 1) estimate the initial noise level of the image sample; 2) obtain the (k + 1)th noise threshold based on the kth noise level; 3) select weak texture regions according to the (k + 1)th noise threshold; 4) estimate the (k + 1)th noise level for each weak texture region; 5) determine whether the (k + 1)th noise level is stable. If it is not stable, return to step 1). If it is stable, the current noise level is the final noise level of the image sample.

[0012] Furthermore, the semantic segmentation model based on deep learning performs a comparison and analysis on the reference image and the image to be detected.

[0013] The beneficial effects of the present invention are as follows: Aiming at the problems of missed reports and false reports in the existing defect recognition method based on image processing, when performing defect detection, the image to be detected is first input into the positive sample detection model to determine whether the image to be detected is abnormal. The image to be detected determined to be abnormal is then input into the negative detection model for defect recognition to identify the specific defect type. The positive sample detection process and the negative sample detection process are carried out in sequence. On the basis of reducing the recognition complexity, the missed report and false report rates are reduced, the reliability of the detection result is improved, the number of pictures for manual review is reduced, and the operation and maintenance and management costs of the system are effectively reduced.

[0014] The defect recognition device based on image processing provided by the present invention to solve the above technical problems includes a processor and a memory. The processor is used to execute the computer program stored in the memory, and the computer program is used to implement the above-mentioned defect recognition method based on image processing.

[0015] The beneficial effects of the present invention are as follows: Aiming at the problems of missed reports and false reports in the existing defect recognition device based on image processing, when performing defect detection, the image to be detected is first input into the positive sample detection model to determine whether the image to be detected is abnormal. The image to be detected determined to be abnormal is then input into the negative detection model for defect recognition to identify the specific defect type. The positive sample detection process and the negative sample detection process are carried out in sequence. On the basis of reducing the recognition complexity, the missed report and false report rates are reduced, the reliability of the detection result is improved, the number of pictures for manual review is reduced, and the operation and maintenance and management costs of the system are effectively reduced.

[0016] The substation intelligent inspection system provided by the present invention to solve the above technical problems includes an image acquisition device and a defect recognition device. The image acquisition device is connected to the defect recognition device. The image acquisition device inputs the acquired image samples into the defect recognition device, and the defect recognition device adopts the above-mentioned defect recognition device based on image processing.

[0017] The beneficial effects of the present invention are as follows: Aiming at the problems of missed reports and false reports in the defect recognition of the existing intelligent substation patrol system, when performing defect detection, the image to be detected is first input into the positive sample detection model to determine whether the image to be detected is abnormal. The image to be detected determined to be abnormal is then input into the negative detection model for defect recognition to identify the specific defect type. By performing the positive sample detection process and the negative sample detection process in sequence, on the basis of reducing the recognition complexity, the missed report and false report rates are reduced, the reliability of the detection result is improved, and the number of pictures for manual review is reduced, effectively reducing the operation and maintenance and management costs of the system. Description of the Drawings

[0018] Figure 1 is the principle block diagram of the defect recognition method based on image processing according to an embodiment of the present invention; Figure 2 is the flowchart of the positive sample detection method according to an embodiment of the present invention; Figure 3 is the flowchart of the negative sample detection method according to an embodiment of the present invention; Figure 4 is the schematic diagram of the regional fusion framework algorithm according to an embodiment of the present invention; Figure 5 is the schematic diagram of the classification method according to an embodiment of the present invention; Figure 6 is the flowchart of the image sample cleaning according to an embodiment of the present invention; Figure 7 is the flowchart of the image sample noise estimation according to an embodiment of the present invention; Figure 8 is the flowchart of the blur degree level estimation algorithm for the image sample according to an embodiment of the present invention; Figure 9 is the flowchart of the preset position offset detection for the image sample according to an embodiment of the present invention. Detailed Embodiments

[0019] The following further describes the detailed embodiments of the present invention with reference to the drawings.

[0020] The basic idea of the present invention is: The present application constructs a positive sample detection model and a negative sample detection model. When performing defect detection, the image to be detected is first input into the positive sample detection model to determine whether the image to be detected is abnormal. The image to be detected determined to be abnormal is then input into the negative detection model for defect recognition to identify the specific defect type. By performing the positive sample detection process and the negative sample detection process in sequence, on the basis of reducing the recognition complexity, the probabilities of missed reports and false reports are reduced, and the reliability and accuracy of the detection result are improved.

[0021] Embodiment of the Defect Recognition Method Based on Image Processing Based on the above basic idea, asFigure 1 As shown in the figure, the defect recognition method based on image processing of the present invention includes: inputting the acquired image to be detected into the positive sample detection model to determine whether there are defects in the image to be detected, and inputting the image to be detected with defects into the negative sample detection model to obtain the defect type of the image to be detected; the positive sample detection model is: selecting the positive sample with the highest similarity to the image to be detected from the constructed positive sample database as the reference image, comparing and analyzing the reference image and the image to be detected, and outputting the comparison result; the negative sample detection model is trained through a negative sample set, and the negative sample set is composed of image samples containing different types of defects and the corresponding defect types.

[0022] In terms of defect recognition, the existing types of image recognition algorithms are insufficient and cannot cover all defects. Especially in the substation scenario, the substation environment is complex. Currently, the substation image recognition algorithms mainly include several types of algorithms such as equipment status class, meter recognition class, equipment defect class, environmental risk class, and personnel behavior class. However, the number of truly practical algorithms is only more than 40 types, which cannot cover all defect types in the substation. In this embodiment, through positive sample detection first, the images determined to be abnormal by the positive sample detection algorithm are subjected to intelligent defect recognition. Therefore, compared with directly using the negative sample detection algorithm to recognize the image to be detected, the false alarm rate and missed alarm rate can be reduced, and the detection accuracy can be improved.

[0023] When detecting the inspection points of the status class and meter class, compared with the original positive samples (initial status or readings), the change in the status of the device recognized by image recognition and the change in the reading numbers and pointers are actually a kind of defect (image difference in the detection area) compared with the original positive samples themselves. Through this "defect", the current status or reading of the device can be obtained.

[0024] As a preferred implementation, a semantic segmentation model based on deep learning is used to compare and analyze the reference image and the image to be detected. As Figure 2 shown, the semantic segmentation model based on deep learning includes three parts: preprocessing, deep feature encoding, and deep feature decoding. The positive sample detection process of the semantic segmentation model based on deep learning is: First, perform online learning on the constructed positive sample database, and select the positive sample with the highest similarity to the image to be detected from the constructed positive sample database as the reference image. Second, input the reference image and the image to be detected after preprocessing into the semantic segmentation model based on deep learning for comparison detection (including the process of matching the dense feature points and sparse feature points of the image to be detected and the reference image), and then post-process and filter the feature map output by the model, and use the changed area of the filtered feature map to represent abnormal defects. Therefore, after the positive sample detection of the image to be detected, not only whether there are defects is detected, but also the changed area of the feature map, that is, the abnormal defect area, is output, which is convenient for subsequent negative sample detection.

[0025] Preferably, the negative sample detection model includes a feature extraction network, a candidate region extraction algorithm, a region fusion algorithm, and a classification network. As Figure 3 shown, the negative sample detection process is as follows: Input the image to be detected into the feature extraction network, and at the same time, process the image to be detected successively using the candidate region extraction algorithm and the region fusion algorithm. Input the processing results and the output results of the feature extraction network into the classification network to finally obtain the recognition result.

[0026] As a preferred implementation, in order to improve the negative sample recognition effect, the candidate region extraction and fusion algorithm adopts algorithms such as MCG, Selective Search, Bing, and Edge Box, as Figure 4 shown. Among them, the MCG and Selective Search algorithms start from the bottom-layer segmentation of the image and give candidate regions where objects may exist in the image in an upward merging manner. The Bing algorithm gives candidate regions similar to the target category from the HOG features of the object contour. The Edge Box algorithm starts from image edge detection, clusters and merges different edges to obtain candidate regions of the target category. Finally, the target candidate regions generated by the four different algorithms are merged through non-maximum suppression (NMS), and the candidate region with the highest confidence is retained and input into the classification network for region classification.

[0027] Preferably, the feature extraction network uses a multi-layer feature fusion method to obtain the network features of a given region to improve the object recognition results at different scales. The feature extraction process includes: First, perform a forward calculation using a deep network; Second, perform region feature sampling on the planes of the last few convolutional layers respectively; Third, perform normalization using L2 Norm and then merge them into a feature vector.

[0028] Preferably, fuse the Softmax and SVM classification methods to classify the region features. As Figure 5 shown, during the training process, use Softmax for backpropagation (BP) to learn parameters; during the training process, use SVM for linear multi-classification of the region features; during the recognition process, perform L2 Norm normalization on the pseudo-probabilities obtained by Softmax; during the recognition process, perform L2 Norm normalization on the probabilities obtained by SVM; adopt the Group Lasso regularization algorithm to fuse the probabilities to obtain the final classification result.

[0029] When constructing a positive sample database, it is necessary to collect image samples. When collecting image samples, if the quality of the collected image samples is poor, it cannot meet the requirements of intelligent applications. Especially in the application scenario of intelligent substation patrol, the core of intelligent substation patrol, intelligent operation, and intelligent safety lies in the intelligent analysis of on-site pictures. However, due to various failures caused by long-term use and wear of image acquisition devices such as cameras, the influence of environmental light, the blurred focus of the camera pan-tilt, and the offset of the preset position, the quality of the original pictures collected by the system is poor, affecting the final use effect of intelligent applications. Therefore, when constructing a positive sample database, it is necessary to preprocess the collected image samples before inputting them into the positive sample detection model, that is, image cleaning. The image cleaning stage mainly undertakes two tasks: one is to pre-eliminate samples that do not meet the image recognition requirements such as black screens, colored screens, and preset position offsets in advance, and re-obtain samples to minimize the picture quality problems caused by situations such as the preset position not turning in place and the focus not being completed; the other is to consider the on-site environmental factors and preprocess the pictures for defogging, denoising, removing backlight, and deblurring in advance to improve the quality of image samples.

[0030] Preferably, in this embodiment, when performing image cleaning, image samples with signal loss, as well as image samples whose image clarity, brightness, and preset position offset do not meet the image recognition conditions, are eliminated, and a positive sample database is established with the acquisition device points as units.

[0031] Such as Figure 6 shown, it is necessary to detect in real time whether the image acquisition device of the inspection equipment has a signal loss phenomenon. When a signal loss occurs, a signal loss warning should be issued. The method for judging whether there is a signal loss in the collected image samples is as follows: If the image sample has at least one of the following situations, it is considered that there is a signal loss: 1) The image is blurred due to incorrect focal length of the shooting device or movement of the shooting device; 2) Noise is introduced during the acquisition, transmission, or processing process, and the noise is greater than the noise threshold; 3) The image is compressed and distorted; 4) There is a colored screen or black screen phenomenon.

[0032] Such as Figure 1 shown, the cleanliness detection includes four parts: brightness anomaly detection, image noise level estimation, image color cast anomaly detection, and image blurriness estimation.

[0033] Preferably, the brightness anomaly detection is to obtain the brightness information of a specific image through histogram analysis, calculate the brightness deviation of the image, and when the deviation is greater than the threshold deviation, it is considered that the brightness is abnormal. At this time, the collected image samples are eliminated and not used to construct a positive sample database.

[0034] Such as Figure 6As shown, the patrol inspection system also makes the following judgments on the collected video images: whether the image signal is lost, whether the image is too dark or too bright. When at least one of the situations of image signal loss, image being too dark, or too bright occurs, subsequent defect detection processing is not performed, and the image is recollected. Further, when at least one of the situations of image signal loss, image being too dark, or too bright occurs, not only is subsequent processing not performed, but corresponding alarms are also issued. To improve the accuracy of judgment, the above judgments are made on two consecutive frames of images. The brightness deviation of the image is greatly affected by the actual lighting environment. The patrol inspection system continuously statistics the brightness changes at each moment, adopts a soft threshold method, adaptively sets the brightness deviation threshold according to the time change, so that it matches the brightness standard at the local moment, thereby improving the accuracy of brightness detection.

[0035] When there is no image signal loss, the image being too dark, or too bright, the noise level of the collected image sample is estimated, the blurriness is estimated based on the noise level, and whether there is color cast is judged, and the pan-tilt rotation situation is detected. If the blurriness is too high (higher than the blurriness threshold), the image sample is severely color cast, or the noise level is abnormal due to the rotation of the pan-tilt, the image sample is discarded and not used.

[0036] Since the image being too dark will cause the image to lose color information, it is necessary to first determine whether the video image is too dark before detecting color cast. The abnormal detection of color cast in the image is mainly based on the color information of the specific image, calculates the color deviation of the RGB three channels of the image, and issues a warning when the color deviation is greater than the color deviation threshold (that is, the image sample is severely color cast).

[0037] Clarity detection is mainly used to evaluate the clarity and focal length quality of the image sample. By extracting features such as the spectral characteristics, contrast, entropy, sharpness, high-frequency components, PSNR (peak signal-to-noise ratio), and SSIM (structural similarity index) of the image sample, the clarity of the image sample is judged through the detection model to obtain high-quality pictures.

[0038] The noise level estimation part of the video image mainly estimates the noise level of the image using the inter-frame difference of the video image. For an image sample with a noise level greater than the threshold, it is considered that there is an abnormality in the acquisition process and a warning needs to be issued. Since frame difference is used, it is necessary to first determine whether there is a sudden change in brightness or the rotation of the pan-tilt. Currently, an iterative method is used to estimate the noise, as Figure 7 shown, the noise level estimation method is: 1) Generate a covariance matrix for all parts of the input image sample to estimate the initial noise level of the image sample σ n (0) ; 2) Based on the k th noise level σn (k) , obtain the k +1 noise threshold τ (k+1) ; 3) According to the k +1 noise threshold τ (k+1) select the weak texture region W (k+1) ; 4) For each weak texture region W (k+1) estimate the k +1 noise level σ n (k+1) ; 5) Determine whether the k +1 noise level σ n (k+1) is stable. If it is not stable, return to step 1). If it is stable, then use the k +1 noise level σ n (k+1) as the final noise level σ n .

[0039] Furthermore, as Figure 8 shown, the blur estimation method is as follows: 1) Estimate the noise level of the image sample; 2) Perform edge detection according to the noise level; 3) Perform multiple re-blurring to obtain the re-blurred image; 4) Obtain the blur degree according to the re-blurred image; 5) Obtain the final blur degree according to the value of the blur degree.

[0040] Image blur is a relatively subjective concept. However, for objective evaluation and processing, it needs to be quantified. Re-blurring can help establish a quantification standard. By blurring the original image to different degrees, a series of image samples with different blur degrees can be generated. These samples can be used as training data or reference standards for training a blur detection model or calibrating a blur quantification index. The re-blurring method can use Gaussian blur. By adjusting the blur parameters (such as the standard deviation of the Gaussian kernel), a series of image samples with different blur degrees can be generated. Preferably, when based on the gradient magnitude and gradient height, according to the gradient magnitude at different blur scales, the first derivative is used to further optimize the blur detection method.

[0041] Furthermore, as Figure 9As shown in the figure, the preset position offset detection process includes: obtaining the original reference preset position image and the image pair to be rectified; calculating and screening the feature descriptors and feature points of the preset position image and the image to be rectified respectively based on the sparse feature points and the SIFT matching algorithm module; screening the sparse feature point pairs to determine whether the number of sparse feature point pairs is greater than the minimum threshold number K , if it is greater, a homography matrix is established according to the calculated sparse feature points; according to the homography matrix, calculate x the offset parameters (image offset pixels) in the y direction and the K direction; finally, compare the offset degrees of the image to be rectified in the two directions to determine whether the offset of the image to be rectified in this direction is greater than the threshold range set by the user; if the offset is greater than the threshold range set by the user, return that the image to be detected is offset and calculate the offset amount, otherwise return that the image to be detected is not offset. If the number of sparse feature point pairs is less than or equal to the minimum threshold number M , then for the original reference preset position image and the image pair to be rectified, calculate the feature descriptors and feature point pairs of the image pair based on the dense feature points and the deep learning algorithm D2-Net module, screen the dense feature point pairs, and determine whether the number of dense feature point pairs is greater than the threshold number x , if it is greater, a homography matrix is established according to the calculated dense feature points; according to the homography matrix, calculate y the offset parameters (image offset pixels) in the M direction and the

[0042] direction; finally, compare the offset degrees of the image to be rectified in the two directions to determine whether the offset of the image to be rectified in this direction is greater than the threshold range set by the user; if the offset is greater than the threshold range set by the user, return that the image to be detected is offset and calculate the offset amount, otherwise return that the image to be detected is not offset. If the number of dense feature point pairs is less than or equal to the threshold number

[0043] Embodiment of the defect recognition device based on image processing Based on the above basic idea, the defect recognition device based on image processing of the present invention includes a processor and a memory. The processor is used to execute the computer program stored in the memory, and the computer program is used to implement the above-mentioned defect recognition method based on image processing. For this method, refer to the embodiment of the defect recognition method based on image processing above, and details are not described here again

[0044] Embodiment of the intelligent substation inspection system The system includes an image acquisition device and a defect recognition device. The image acquisition device is connected to the defect recognition device. The image acquisition device inputs the acquired image samples into the defect recognition device. The defect recognition device is a defect recognition device based on image processing. For this device, refer to the embodiments of the defect recognition device based on image processing above, and no detailed description will be given here.

[0045] Preferably, the image acquisition device uses devices such as high-definition cameras, intelligent inspection robots, and / or drones. The image acquisition device collects photos and video images under different lighting conditions and shooting angles in the substation as image samples, and then processes them using the image cleaning method in the embodiments of the above-mentioned defect recognition method based on image processing, so as to construct a positive sample database.

[0046] The advantages of the substation intelligent inspection system in this embodiment are as follows: Aiming at problems such as missed reports and false alarms in the substation remote intelligent inspection system, samples that do not meet the requirements of image recognition, such as black screens, flower screens, and preset position offsets, during the image acquisition process are removed through image cleaning, and preprocessing such as defogging, denoising, removing backlight, and deblurring is performed on the pictures to improve the quality of the image samples. Through the positive and negative sample recognition method, without missed reports, different regions of the inspection pictures and the bottom map most similar to the normal samples are marked, and specific defects are identified to achieve real-time monitoring of status type and meter type points. Therefore, the substation remote intelligent inspection system in this embodiment improves the sample quality from aspects such as image sample acquisition, detection, and image quality improvement, uses the positive and negative sample image processing method to improve the detection accuracy, realizes the purpose of reducing the missed report rate and false alarm rate of the intelligent inspection system and reducing the number of pictures for manual review, can effectively reduce the operation and maintenance and management costs of system users, and can create considerable economic benefits.

Claims

1. A defect recognition method based on image processing, characterized in that: The method comprises: inputting the acquired image to be detected into a positive sample detection model to determine whether the image to be detected has defects, and inputting the image to be detected with defects into a negative sample detection model to obtain the defect type of the image to be detected; the positive sample detection model comprises: selecting a positive sample with the greatest similarity to the image to be detected from a constructed positive sample database as a reference image, performing a comparison analysis between the reference image and the image to be detected, and outputting a comparison result; The negative sample detection model is obtained by training a negative sample set, where the negative sample set consists of image samples containing different types of defects and corresponding defect types.

2. The defect recognition method based on image processing according to claim 1, characterized in that: The method for constructing the positive sample database is as follows: collecting image samples under different lighting conditions and shooting angles in application scenarios, performing image cleaning on the image samples, eliminating image samples with signal loss, or image samples whose image clarity, brightness or preset position offset do not meet image recognition conditions, and establishing a positive sample database with image acquisition device points as units.

3. The defect recognition method based on image processing according to claim 2, characterized in that: When the image sample has at least one of the following conditions, it is considered that there is signal loss: 1) the image is blurred due to incorrect focus of the camera or movement of the camera; 2) noise is introduced during the acquisition, transmission or processing process, and the noise is greater than the noise threshold; 3) Image compression distortion; 4) The screen is distorted or black.

4. The defect recognition method based on image processing according to claim 2, characterized in that: Image clarity detection includes four parts: brightness anomaly detection, image noise level estimation, image color cast anomaly detection and image blur estimation.

5. The defect recognition method based on image processing according to claim 1, characterized in that: The negative sample detection model includes a feature extraction network, a candidate region extraction and fusion algorithm, and a classification network.

6. The defect recognition method based on image processing according to claim 5, characterized in that: The candidate region extraction and fusion algorithms include four algorithms: MCG, Selective Search, Bing and Edge Box. The four algorithms are used to generate target candidate regions respectively, and each target candidate region is merged through non-maximum suppression to retain the candidate region with the highest confidence.

7. The defect recognition method based on image processing according to claim 4, characterized in that: The steps of estimating the image noise level include: 1) estimating the initial noise level of the image sample; 2) obtaining the k+1th noise threshold based on the kth noise level; 3) selecting a weak texture area according to the k+1th noise threshold; 4) estimating the k+1th noise level for each weak texture area; 5) determining whether the k+1th noise level is stable, and if not, returning to step 1); if it is stable, the current noise level is the final noise level of the image sample.

8. The defect recognition method based on image processing according to claim 1, characterized in that: The semantic segmentation model based on deep learning is used to compare and analyze the reference image and the image to be detected.

9. A defect recognition device based on image processing, characterized in that: The device includes a processor and a memory, wherein the processor is used to execute a computer program stored in the memory, and the computer program is used to implement the image processing-based defect recognition method described in any one of claims 1 to 8.

10. A substation intelligent inspection system, characterized in that: The system includes an image acquisition device and a defect recognition device. The image acquisition device is connected to the defect recognition device. The image acquisition device inputs the acquired image samples into the defect recognition device. The defect recognition device adopts the image processing-based defect recognition device described in claim 9.

Images