Basin-type insulator infrared thermal imaging defect detection system and method
By combining infrared thermal imaging technology and the U-net model of efficient attention module, defect detection of basin insulators is solved, and the problems of inaccurate defect detection and blurred edges in the existing technology are achieved, and high-precision defect segmentation and diameter are achieved Test.
Patent Information
- Application Number
- CN202411451293.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-17
- Publication Date
- 2025-02-14
AI Technical Summary
The existing infrared thermal imaging technology has problems such as blurred image edges, high leakage detection rate and inaccurate detection in basin insulator defect detection, making it difficult to achieve accurate defect positioning and dimension calibration.
Infrared thermal imaging technology is used to collect infrared thermal images of basin insulators, and image processing is performed by combining the U-net model of the efficient attention module, image features are extracted and upsampled operations are performed. The upsampled feature map is connected to the jump connection. The feature maps of different resolutions in the input module are fused to generate segmented maps of insulator defects.
It realizes high-precision segmentation of basin insulator defects, can accurately detect defect diameters, and solves the problems of insufficient defect detection accuracy and blurred internal defect boundaries in traditional methods.
Smart Images
Figure CN119444680A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of infrared thermal imaging detection technology, and in particular to an infrared thermal imaging defect detection system and method for a pot-type insulator. Background Art
[0002] In modern power systems, basin insulators, as key components of gas-insulated switchgear (GIS), insulate and support the GIS's central conductor. Their insulation performance directly impacts the reliability of GIS equipment. In recent years, a growing number of researchers have focused on insulator defect detection, employing pulsed infrared thermal wave nondestructive testing (NDT) techniques and models such as YoLov5 and YoLov8. However, infrared detection methods often miss detection and are inaccurate due to blurred edges in infrared thermal images of internal insulator defects.
[0003] Infrared thermal imaging technology focuses the infrared radiation of an object onto an infrared detector, which then converts the radiation signals of varying strengths into corresponding electrical signals, which are then processed and converted into visual infrared thermal images, thereby extending the visual range of the human eye to the invisible infrared region.
[0004] The image processing algorithm of the improved U-net insulator defect detection method is based on the U-net network model and is improved by integrating the efficient attention module. It performs image processing on a set of 24-bit image information related to the defect location information obtained from the infrared thermal image, intuitively expresses the defect location information contained in the image, and thus calibrates the diameter of the defect.
[0005] The basin-type insulator infrared thermal imaging defect detection system built using infrared thermal imagers and deep learning technology can detect defects in insulator samples and calibrate the defect diameter.
[0006] Infrared thermal imaging technology is a non-destructive testing technology with intuitive results, wide applicability, and ease of qualitative and quantitative analysis. Deep learning technology processes images with the characteristics of fast speed, high accuracy, and low cost compared to manual inspection. It has achieved certain achievements in defect detection and is favored by more and more researchers.
[0007] Infrared thermal wave nondestructive testing technology is a new type of nondestructive testing technology that has developed rapidly and been widely used in the past two decades. It has the characteristics of fast testing speed, high detection accuracy, non-contact measurement, intuitive results, wide range of applications, and easy qualitative and quantitative analysis.
[0008] Deep learning technology is one of the most popular and researched branches of artificial intelligence today. With the advent of the big data era, the vast amounts of data required for deep learning models are now readily available. Furthermore, the continuous improvement of computer hardware performance has provided powerful computing power for training deep learning algorithms. Deep learning has a wide range of applications in image processing and understanding. For example, in the field of face recognition, it can help improve tasks such as face detection, feature extraction, and recognition. It boasts strong learning capabilities, wide coverage, good adaptability, and excellent portability.
[0009] In summary, infrared thermal imaging technology has many advantages, such as being unaffected by external temperature fields, displaying very intuitive images, having a relatively fast response speed, and good concealment. The U-net model has many advantages, such as supporting small sample data training models, having high segmentation accuracy, and being fast. However, the defect images identified by infrared thermal imagers are relatively blurry and difficult to identify. To address this issue, this patent uses infrared thermal imaging technology to collect infrared thermal images of pot-type insulators to construct an insulator sample dataset. The U-net model combined with an efficient attention module is used to process the insulator samples, thereby improving the feature expression capability and achieving high-precision segmentation of defects, facilitating the detection of defect diameters and completing the detection of pot-type insulator defects. Summary of the Invention
[0010] In view of the above-mentioned problems, the present invention is proposed.
[0011] The technical problem addressed by this invention is that existing infrared thermal imaging technology for detecting defects in pot-type insulators suffers from image blurring, high rates of missed detection, and inaccurate detection, making it difficult to accurately locate and calibrate defects. Furthermore, traditional manual inspection methods are time-consuming, labor-intensive, and suffer from poor consistency, impacting inspection efficiency and quality. Furthermore, existing technologies typically rely on large sample sizes and exhibit limitations in acquiring high-quality small sample sizes, limiting the accuracy and practicality of the inspection system.
[0012] To solve the above technical problems, the present invention provides the following technical solutions: a basin-type insulator infrared thermal imaging defect detection system, comprising:
[0013] An input module and a processing module; the input module is used to receive infrared thermal imaging images of pot-type insulators and gradually generate image features through multiple convolution blocks; the processing module performs upsampling operations and fuses the upsampled feature maps with the corresponding feature maps of different resolutions in the input module layer by layer through jump connections to generate a segmentation map of insulator defects.
[0014] As a preferred solution of the basin insulator infrared thermal imaging defect detection system of the present invention, wherein: each convolution block of the input module includes performing two convolution operations on the input image, sliding the image through the convolution kernel and performing element-wise multiplication and summation operations to extract image features;
[0015] After the convolution operation, the convolution results are batch normalized by calculating the mean and variance.
[0016] As a preferred solution of the pot-type insulator infrared thermal imaging defect detection system described in the present invention, the input module uses global average pooling to reduce the dimension of image features through the ECA module after batch normalization, and learns the weight of each channel through a one-dimensional convolution operation.
[0017] As a preferred solution of the basin insulator infrared thermal imaging defect detection system of the present invention, wherein: the input module performs a maximum pooling operation after the feature weighting module;
[0018] The maximum pooling operation downsamples the image features through a fixed pooling step size to obtain reduced image features.
[0019] As a preferred solution of the basin-type insulator infrared thermal imaging defect detection system described in the present invention, the processing module includes multiple upsampling units, each of which restores image features layer by layer through a transposed convolution operation and gradually increases the resolution of the image features to approach the original image size.
[0020] As a preferred solution of the infrared thermal imaging defect detection system for pot-type insulators of the present invention, wherein: the processing module splices the upsampled image features with the corresponding image features of different resolutions in the input module through a jump connection after each upsampling unit;
[0021] Multiple convolution operations are performed on the spliced image features to generate a defect segmentation map of the same size as the original input image through 1×1 convolution.
[0022] As a preferred solution of the infrared thermal imaging defect detection system for pot-type insulators described in the present invention, the defect position in the defect segmentation map is identified and calibrated to obtain the circumference of the defect, and the defect diameter is obtained by calculating the circumference formula to complete the pot-type insulator defect diameter detection.
[0023] Another object of the present invention is to provide a method for detecting defects in pot-type insulators using infrared thermal imaging, which can accurately detect the location and size of defects in pot-type insulators through multi-level feature extraction and segmentation, thereby solving the problems of insufficient defect detection accuracy and blurred internal defect boundaries in traditional methods.
[0024] In order to solve the above technical problems, the present invention provides the following technical solutions: a method for detecting defects of pot-type insulators by infrared thermal imaging, comprising: receiving infrared thermal imaging images of pot-type insulators through an input module, and inputting the images into multiple convolution blocks; performing two convolution operations on the input image in each convolution block of the input module, and sliding the convolution kernel on the image to extract image features; after each convolution, performing batch normalization on the convolution results, and standardizing the data distribution by calculating the mean and variance; after batch normalization, using global average pooling to reduce the dimension of the image features through the ECA module, and learning the weight of each channel through a one-dimensional convolution operation to form a weighted feature map; after the feature weighting module, performing a maximum pooling operation, and reducing the weighted image features through a fixed pooling step size. The reduced image features are obtained by sampling; the reduced image features are passed to the processing module, which contains multiple upsampling units. The resolution of the image features is restored layer by layer through the transposed convolution operation, making it gradually close to the size of the original input image; after each upsampling unit, the upsampled image features are spliced with the corresponding image features of different resolutions in the input module through jump connections, and the spliced feature map is used as the input of the next convolution operation; multiple convolution operations are performed on the spliced image features, and finally a defect segmentation map of the same size as the original input image is generated through 1×1 convolution; the defect segmentation map is analyzed, the defect location is identified and calibrated; the circumference of the defect is calculated based on the calibration result, and the defect diameter is calculated based on the circumference, thereby completing the defect diameter detection of pot insulators.
[0025] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the above-mentioned method for detecting defects of pot-type insulators by infrared thermal imaging are implemented.
[0026] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for detecting defects of pot-type insulators by infrared thermal imaging.
[0027] Beneficial effects of the present invention: The infrared thermal imaging defect detection method for pot-type insulators provided by the present invention is applied to defect detection of pot-type insulators. Infrared thermal imaging is used in an infrared thermal imaging system for non-destructive testing. The method does not require contact with the target object, provides temperature information of the target object, does not disturb the target temperature, and the temperature distribution is intuitively visible.
[0028] The U-net model in the ECA-Unet model of the present invention can classify each pixel, thereby achieving high segmentation accuracy and having good small sample learning ability. Its symmetrical encoder-decoder structure reduces the spatial size of the feature map when extracting image features, while the decoder gradually restores the spatial size of the image and refines the segmentation results, helping to recover image details while maintaining sufficient context information. The feature map in the encoder and the feature map in the decoder are combined through jump connections, which helps to retain more spatial information in the decoder, thereby improving segmentation accuracy.
[0029] The ECA efficient attention module in the ECA-Unet model of the present invention captures the dependencies between channels through one-dimensional convolution, avoiding the complex dimensionality reduction and dimensionality increase process, thereby achieving high efficiency and lightweight characteristics;
[0030] The present invention simultaneously calculates the defect diameter through the perimeter, and the calculation is simple, direct and rapid, and can also be performed on irregular shapes. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.
[0032] Figure 1 This is an overall structural diagram of an infrared thermal imaging defect detection system for pot-type insulators provided by one embodiment of the present invention.
[0033] Figure 2 A detection structure diagram of an infrared thermal imaging defect detection system for pot-type insulators provided by one embodiment of the present invention.
[0034] Figure 3 A detection flow chart of an infrared thermal imaging defect detection system for pot-type insulators provided in one embodiment of the present invention.
[0035] Figure 4 A schematic diagram of a pot-type insulator sample preparation process for a pot-type insulator infrared thermal imaging defect detection system provided by an embodiment of the present invention.
[0036] Figure 5 A schematic diagram of the structure of an insulator defect detection optimization algorithm of an infrared thermal imaging defect detection system for pot-type insulators provided by one embodiment of the present invention.
[0037] Figure 6A schematic diagram of an efficient attention module of an infrared thermal imaging defect detection system for pot-type insulators provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0038] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.
[0039] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.
[0040] Example 1
[0041] Reference Figure 1 , as one embodiment of the present invention, provides a method for detecting defects of pot-type insulators using infrared thermal imaging, comprising:
[0042] Input module 100 and processing module 200;
[0043] The input module 100 is used to receive infrared thermal imaging images of pot-type insulators and gradually generate image features through multiple convolution blocks;
[0044] The processing module 200 performs an upsampling operation and fuses the upsampled feature map with the corresponding feature maps of different resolutions in the input module layer by layer through skip connections to generate a segmentation map of the insulator defect.
[0045] In this embodiment of the present invention, the input module 100 is implemented by an encoder. In the encoder part, a three-channel image of size 480×640 is input, and the image is convolved twice. A 3×3 convolution kernel is slid on the image and element-wise multiplication and summation operations are performed to extract the features of the image. The size of the image after convolution is 638×476.
[0046] Batch normalization is then performed. By calculating the mean and variance of small batches of data, these statistics are then used to normalize the data and normalize the output of the intermediate layers of the neural network to reduce internal covariate shift. This ensures that the distribution of input data at each layer of the network remains relatively stable. Finally, the expressive power of the network is restored through learnable parameters.
[0047] After this operation, the ECA module is added. First, global average pooling is used to reduce the dimension of the feature map. Then, the weight of each channel is learned through a one-dimensional convolution operation. These weights are then applied to each channel of the input feature map to achieve a weighted combination of different channels.
[0048] The obtained feature map is subjected to a maximum pooling operation with a pooling step of 2 to reduce the spatial size of the feature map while retaining important feature information, reducing the number of model parameters and the amount of computation. The resulting feature map size is 319×238, completing the operation of the first convolutional block.
[0049] The following convolution blocks are the same as the above steps
[0050] In the second convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 317×236;
[0051] Then pass through the batch normalization layer, the size does not change;
[0052] The size after max pooling (2×2, stride 2) is 159×118;
[0053] In the third convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 157×116;
[0054] Then pass through the batch normalization layer, the size does not change;
[0055] The size after max pooling (2×2, stride 2) is 79×58;
[0056] In the fourth convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 77×56;
[0057] Then pass through the batch normalization layer, the size does not change;
[0058] The size after max pooling (2×2, stride 2) is 39×28;
[0059] In the fifth convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 37×26.
[0060] In the present invention, the processing module 200 is implemented by a decoder, and the feature map obtained at this time enters the decoder part:
[0061] The feature map obtained after the fifth convolution block is upsampled using a 2×2 transposed convolution with a stride of 2 to restore the size of the feature map of this layer to the size of the feature map of the previous layer, which is 74×52.
[0062] Through skip connection and concatenation with the feature map of the fourth convolution block, two 3×3 convolutions are performed, and the size after convolution is 75×53;
[0063] The second upsampling (2×2 transposed convolution, stride 2) results in a size of 150×106. This is then concatenated with the feature map of the third convolutional block, and then undergoes two 3×3 convolutions to a size of 151×108.
[0064] The third upsampling (2×2 transposed convolution, stride 2), the upsampling size is 302×212
[0065] It is concatenated with the feature map of the second convolutional block, and then undergoes two 3×3 convolutions to a size of 303×214;
[0066] The fourth upsampling (2×2 transposed convolution, stride 2), the upsampling size is 606×428
[0067] It is concatenated with the feature map of the first convolutional block, and then undergoes two 3×3 convolutions to a size of 607×429;
[0068] Finally, after a 1×1 convolution, the feature map is mapped to the number of target categories, and the final output size is 640×480, and the feature map is restored to the original image size.
[0069] Each convolution block of the input module 100 includes performing two convolution operations on the input image, sliding the image through the convolution kernel and performing element-wise multiplication and summation operations to extract image features;
[0070] After the convolution operation, the convolution results are batch normalized by calculating the mean and variance.
[0071] The input module 100 uses global average pooling to reduce the dimension of image features through the ECA module after batch normalization, and learns the weight of each channel through a one-dimensional convolution operation.
[0072] The input module 100 performs a maximum pooling operation after the feature weighting module;
[0073] The maximum pooling operation downsamples the image features through a fixed pooling step size to obtain reduced image features.
[0074] The processing module 200 includes a plurality of upsampling units, each of which restores image features layer by layer through a transposed convolution operation and gradually increases the resolution of the image features to approach the original image size.
[0075] The processing module 200 concatenates the upsampled image features with the corresponding image features of different resolutions in the input module 100 through a skip connection after each upsampling unit;
[0076] Multiple convolution operations are performed on the spliced image features to generate a defect segmentation map of the same size as the original input image through 1×1 convolution.
[0077] At this time, the segmentation map of the defect is obtained. By identifying the defect position and calibrating it, the circumference of the defect is obtained. The defect diameter is obtained by calculating the circumference formula, and the defect diameter detection of the pot insulator is completed.
[0078] Example 2
[0079] Reference Figure 2 , as one embodiment of the present invention, provides a basin-type insulator infrared thermal imaging defect detection device, including an infrared thermal imaging system and an improved U-net insulator defect detection optimization algorithm;
[0080] Infrared thermal imaging system, including pot insulator specimen, infrared thermal imager and other devices;
[0081] Improve the U-net insulator defect detection optimization algorithm, including: fusing the efficient attention module with U-net to obtain ECA-Unet (Efficient Convolutional Attention-Unet), and calibrating the diameter of the defect segmentation map after segmentation;
[0082] The power amplifier is connected to the halogen lamp and the signal generator to enhance the power of the halogen lamp and improve the signal strength so that the signal will not be lost due to attenuation during transmission.
[0083] The signal generator is connected to the control system to monitor the output signal in real time;
[0084] Use a 5000W halogen lamp to irradiate the pot-type insulator sample at a distance of 13 cm. After irradiation for 6 seconds, turn off the halogen lamp and slowly cool it down through a cooling device.
[0085] The infrared lens in the infrared thermal imager captures the temperature of the sample, images it through the information processing unit and image processing unit, and connects to the control system via a USB cable to collect infrared thermal images on the control system side;
[0086] The infrared thermal image is labeled, then learned through the ECA-Unet model, and finally segmented to obtain its segmentation map. The defect diameter is then obtained by calculating the circumference formula.
[0087] The present invention is mainly used for defect detection of pot-type insulators. Infrared thermal imaging is used in an infrared thermal imaging system for non-destructive testing. It does not require contact with the target object, provides temperature information of the target object, does not disturb the target temperature, and the temperature distribution is intuitively visible.
[0088] The U-net model in the ECA-Unet model of the present invention can classify each pixel, thereby achieving high segmentation accuracy and having good small sample learning ability. Its symmetrical encoder-decoder structure reduces the spatial size of the feature map when extracting image features, while the decoder gradually restores the spatial size of the image and refines the segmentation results, helping to recover image details while maintaining sufficient context information. The feature map in the encoder and the feature map in the decoder are combined through jump connections, which helps to retain more spatial information in the decoder, thereby improving segmentation accuracy.
[0089] The ECA efficient attention module in the ECA-Unet model of the present invention captures the dependencies between channels through one-dimensional convolution, avoiding the complex dimensionality reduction and dimensionality increase process, thereby achieving high efficiency and lightweight characteristics;
[0090] The present invention simultaneously calculates the defect diameter through the perimeter, and the calculation is simple, direct and rapid, and can also be performed on irregular shapes.
[0091] Example 3
[0092] Reference Figure 3-Figure 4 , as one embodiment of the present invention, provides an infrared thermal imaging system, including a pot-type insulator sample, an infrared thermal imager and other devices;
[0093] The pot-type insulator specimens were prepared by stirring bisphenol A epoxy resin, allowing it to stand, pouring it into a mold, injecting air bubbles, curing it, and demoulding it.
[0094] The infrared thermal imager includes an infrared lens, an information processing unit, and an image processing unit;
[0095] Other devices include halogen lamps, control systems, cooling devices, power amplifiers, and signal generators;
[0096] The infrared thermal imager uses the A655sc scientific research thermal imager produced by FLIR. It uses an optical system to collect infrared radiation from the target and surrounding background, and uses detection elements to convert it into electrical signals for imaging.
[0097] The output end of the power amplifier is connected to the input end of the halogen lamp to amplify the power of the halogen lamp;
[0098] The output of the power amplifier is connected to the input of the signal generator to increase the signal strength so that the signal will not lose information due to attenuation during transmission;
[0099] The signal generator is connected to the control system to monitor the output signal in real time;
[0100] above Figure 3The pot insulator pattern mentioned in the reference is made of the same material as the pot insulator. In order to better observe the defects of the pot insulator, an algorithm is used to learn its features and obtain the segmentation map of the pot insulator. Figure 4 ,like Figure 4 As shown in the figure, the sample preparation process of pot insulator is as follows:
[0101] The pot insulator sample is made of industrial grade bisphenol A epoxy resin with the epoxy value of 0.26 mol / 100g and is poured into a measuring cup. The sample is stirred with a stirring rod until it is completely transparent and there is no
[0102] To prevent the wire drawing phenomenon, pour it into a round airbag silicone mold after letting it stand for 30 minutes. Move the mold into a constant temperature box at 25°C. Keep the sample in a constant temperature environment at 25°C for 30 minutes, inject bubbles into it with a syringe, and control their size and distribution. Move the sample mold after bubble injection into a constant temperature box at 25°C and wait for 10 to 24 hours. After the sample is completely cured, take the sample out of the mold to complete the sample preparation.
[0103] Example 4
[0104] Reference Figure 5-Figure 6 , as one embodiment of the present invention, provides a method for detecting defects of pot-type insulators using infrared thermal imaging, comprising:
[0105] like Figure 5 As shown in the figure, the efficient attention module is integrated with U-net to obtain ECA-Unet (EfficientConvolutionalAttention-Unet, and the diameter of the defect segmentation map after segmentation is calibrated;
[0106] The specific steps to merge the two modules are as follows:
[0107] In the encoder part, a 480×640 three-channel image is input and convolved twice. A 3×3 convolution kernel slides over the image and performs element-wise multiplication and summation operations to extract the image features. The size of the image after convolution is 638×476.
[0108] Batch normalization is then performed. By calculating the mean and variance of small batches of data, these statistics are then used to normalize the data and normalize the output of the intermediate layers of the neural network to reduce internal covariate shift. This ensures that the distribution of input data at each layer of the network remains relatively stable. Finally, the expressive power of the network is restored through learnable parameters.
[0109] After this operation, the ECA module is added. First, global average pooling is used to reduce the dimension of the feature map. Then, the weight of each channel is learned through a one-dimensional convolution operation. These weights are then applied to each channel of the input feature map to achieve a weighted combination of different channels.
[0110] The obtained feature map is subjected to a maximum pooling operation with a pooling step of 2 to reduce the spatial size of the feature map while retaining important feature information, reducing the number of model parameters and the amount of computation. The resulting feature map size is 319×238, completing the operation of the first convolutional block.
[0111] In the second convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 317×236;
[0112] Then pass through the batch normalization layer, the size does not change;
[0113] The size after max pooling (2×2, stride 2) is 159×118;
[0114] In the third convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 157×116;
[0115] Then pass through the batch normalization layer, the size does not change;
[0116] The size after max pooling (2×2, stride 2) is 79×58;
[0117] In the fourth convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 77×56;
[0118] Then pass through the batch normalization layer, the size does not change;
[0119] The size after max pooling (2×2, stride 2) is 39×28;
[0120] In the fifth convolution block, the convolution kernel is 3×3, the stride is 1, and the size after convolution is 37×26;
[0121] The feature map obtained at this time enters the decoder part:
[0122] The feature map obtained after the fifth convolution block is upsampled using a 2×2 transposed convolution with a stride of 2 to restore the size of the feature map of this layer to the size of the feature map of the previous layer, which is 74×52.
[0123] Through skip connection and concatenation with the feature map of the fourth convolution block, two 3×3 convolutions are performed, and the size after convolution is 75×53;
[0124] The second upsampling (2×2 transposed convolution, stride 2) results in a size of 150×106. This is then concatenated with the feature map of the third convolutional block, and then undergoes two 3×3 convolutions to a size of 151×108.
[0125] The third upsampling (2×2 transposed convolution, stride 2), the upsampling size is 302×212
[0126] It is concatenated with the feature map of the second convolutional block, and then undergoes two 3×3 convolutions to a size of 303×214;
[0127] The fourth upsampling (2×2 transposed convolution, stride 2), the upsampling size is 606×428
[0128] It is concatenated with the feature map of the first convolutional block, and then undergoes two 3×3 convolutions to a size of 607×429;
[0129] Finally, after a 1×1 convolution, the feature map is mapped to the number of target categories, and the final output size is 640×480, and the feature map is restored to the original image size.
[0130] At this time, the segmentation map of the defect is obtained. By identifying the defect position and calibrating it, the circumference of the defect is obtained. The defect diameter is obtained by calculating the circumference formula, and the defect diameter detection of the pot insulator is completed.
[0131] In the U-net model, after inputting the original image, the model first performs feature extraction and then uses a concatenation operation to fuse low-level feature maps with high-level feature maps. Its completely symmetrical U-shaped structure allows for the addition of both low-resolution and high-resolution information in the resulting image, resulting in a more thorough fusion of high- and low-level features. It also combines the coarse information from downsampling with the high-precision information from upsampling. Furthermore, a fusion operation is used to fill in the underlying information, thereby improving image segmentation accuracy. This network architecture consists of two parts: an encoder and a decoder. The encoder, as the hub of the entire process, primarily extracts features from the input image. To accommodate the transition from higher spatial resolution to lower resolution, the encoder also reduces the image resolution to capture finer details. The decoder then uses an upsampling operation to restore the feature maps to their original size. This technique enlarges these feature maps to the same size as the original input image. Based on these enlarged feature maps, the decoder gradually constructs a segmentation result, identifying individual objects or regions within the image.
[0132] The U-Net algorithm introduces a unique skip connection mechanism. This allows the decoder to directly access feature maps processed by the encoder without re-performing any forward or backpropagation operations. This reduces unnecessary data transmission and processing time, while also improving the efficiency and speed of the model.
[0133] refer to Figure 6 In the efficient attention module, global average pooling is performed on the input feature map without dimensionality reduction, keeping the channel dimension unchanged while compressing the spatial dimension to a dimension of size 1. The last dimension of the aggregated feature is compressed and the dimensions are rearranged. By considering each channel and its k nearest neighbors to capture local cross-channel interaction information, ECA can be effectively implemented through fast 1D convolutions of size k to complete a local cross-channel interaction strategy without dimensionality reduction, where k represents the coverage of local cross-channel interactions, and the interaction coverage k is proportional to the channel dimension. This helps to more effectively conduct channel-to-channel interactions while maintaining inter-channel correlation, thereby improving the network's expressiveness and performance. The output is then passed through a sigmoid activation function to ensure that the output is between 0 and 1. The normalized output is then transformed to restore its shape. Finally, the channel attention weights obtained in the previous step are multiplied by the original input feature map.
[0134] Given an input Through a global average pooling, it becomes In order to adjust the obtained feature map to the shape required by the subsequent convolution operation, it becomes After one-dimensional convolution, the weight of one-dimensional convolution is formula (1):
[0135]
[0136] Among them, Ω i k The set of k adjacent channels represented by is used to implement information interaction between channels through one-dimensional convolution of size k, as shown in formula (2):
[0137] w=σ(C1D k (Y)) (2)
[0138] Where C1D represents one-dimensional convolution, and the size of k is proportional to the channel dimension C, that is, there is a mapping φ between k and C, and the relationship is as shown in formula (3):
[0139] C=φ(k) (3)
[0140] This mapping is set as a simple linear function, that is, φ(k) = (γ*kb), and the channel dimension C is usually set to a power of 2, so the nonlinear mapping relationship is obtained as shown in formula (4):
[0141] C=φ(k)=2 (γ*k-b) (4)
[0142] Given the channel dimension C, the kernel size k can be adaptively obtained by formula (5),
[0143]
[0144] Among them, |t odd represents the odd number closest to t, and γ and b are usually set to 2 and 1.
[0145] Then the output is passed through the sigmoid activation function, and finally the dimension of the standardized output is transformed to restore its shape. Finally, the channel attention weight obtained after convolution is multiplied by the original feature map of the input to obtain the final result.
[0146] Example 5
[0147] An embodiment of the present invention is different from the first four embodiments in that:
[0148] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, 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. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.
[0149] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.
[0150] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.
[0151] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.
[0152] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.
Claims
1. A basin insulator infrared thermal imaging defect detection system, characterized in that: include: An input module (100) and a processing module (200); The input module (100) is used to receive an infrared thermal imaging image of a pot-type insulator and gradually generate image features through a plurality of convolution blocks; The processing module (200) performs an upsampling operation, and fuses the upsampled feature map with the corresponding feature maps of different resolutions in the input module layer by layer through skip connections, thereby generating a segmentation map of insulator defects.
2. The infrared thermal imaging defect detection system for pot-type insulators according to claim 1, characterized in that: Each convolution block of the input module (100) includes performing two convolution operations on the input image, sliding the image through the convolution kernel and performing element multiplication and summation operations to extract image features; After the convolution operation, the convolution results are batch normalized by calculating the mean and variance.
3. The infrared thermal imaging defect detection system for pot-type insulators according to claim 2, characterized in that: The input module (100) uses global average pooling to reduce the dimension of image features through the ECA module after batch normalization, and learns the weight of each channel through a one-dimensional convolution operation.
4. The infrared thermal imaging defect detection system for pot-type insulators according to claim 3, characterized in that: The input module (100) performs a maximum pooling operation after the feature weighting module; The maximum pooling operation downsamples the image features through a fixed pooling step size to obtain reduced image features.
5. The pot-type insulator infrared thermal imaging defect detection system according to claim 4, characterized in that: The processing module (200) comprises a plurality of up-sampling units, each of which restores the image features layer by layer through a transposed convolution operation, and gradually increases the resolution of the image features to approach the original image size.
6. The pot-type insulator infrared thermal imaging defect detection system according to claim 5, characterized in that: The processing module (200) splices the upsampled image features with the corresponding image features of different resolutions in the input module (100) through a skip connection after each upsampling unit; The spliced image features are subjected to multiple convolution operations, and a defect segmentation map of the same size as the original input image is generated through 1×1 convolution.
7. The infrared thermal imaging defect detection system for pot-type insulators according to claim 6, characterized in that: By identifying the defect position in the defect segmentation map, the circumference of the defect is obtained by calibration, and the defect diameter is obtained by calculating the circumference formula to complete the defect diameter detection of the pot insulator.
8. A method using the pot insulator infrared thermal imaging defect detection system according to any one of claims 1 to 7, characterized in that: include: receiving an infrared thermal imaging image of a pot-type insulator through an input module, and inputting the image into a plurality of convolution blocks; In each convolution block of the input module, the input image is convolved twice, and the convolution kernel slides on the image to extract image features; After each convolution, batch normalization is performed on the convolution results to standardize the data distribution by calculating the mean and variance; After batch normalization, the image features are reduced in dimension using global average pooling through the ECA module, and the weight of each channel is learned through a one-dimensional convolution operation to form a weighted feature map; After the feature weighting module, the maximum pooling operation is performed to downsample the weighted image features through a fixed pooling step size to obtain the reduced image features; The reduced image features are passed to the processing module, which contains multiple upsampling units. The resolution of the image features is restored layer by layer through transposed convolution operations, making it gradually approach the size of the original input image. After each upsampling unit, the upsampled image features are concatenated with the corresponding image features of different resolutions in the input module through skip connections, and the concatenated feature map is used as the input for the next convolution operation; Perform multiple convolution operations on the spliced image features, and finally generate a defect segmentation map of the same size as the original input image through 1×1 convolution; Analyze the defect segmentation map, identify the defect location and calibrate it; The circumference of the defect is calculated based on the calibration results, and the defect diameter is calculated through the circumference, thereby completing the defect diameter detection of the pot insulator.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method for detecting defects of pot-type insulators by infrared thermal imaging are implemented as described in claim 7.
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 steps of the method for detecting defects of pot-type insulators by infrared thermal imaging are implemented as described in claim 7.