A network and method for monitoring and removing reflection interference of a panel pressure plate of a UHV converter station

Abstract

The invention discloses a network and method for monitoring and removing reflection interference of a pressure plate of a UHV converter station cabinet, comprising a plurality of sequentially cascaded encoders and a plurality of sequentially cascaded decoders, wherein the encoder comprises two sequentially connected residual blocks and a WRNL block, the decoder comprises a sequentially connected convolution layer, two residual blocks and a WRNL block, the convolution layer is connected to an SE block, the output result of the previous stage fused at the input end of the current decoder and the output result of each encoder are input to the SE block in the current decoder, the SE block adjusts the number of channels through the convolution layer and then inputs the result to the residual block in the current decoder, an image with reflection interference is input to the first-stage encoder, and the last-stage decoder outputs an image with the reflection interference removed; the invention has the advantages of solving the problem that the surface glass reflection of the cabinet in the panoramic monitoring process of the UHV DC converter station equipment makes it difficult to monitor part of the collected image.

Description

Technical Field

[0001] The present invention relates to the field of lightweight artificial intelligence edge-side distributed data processing for ultra-high voltage direct current protection systems, and more specifically to an ultra-high voltage converter station panel pressure plate monitoring and de-reflection interference network. Background Art

[0002] Sometimes the protection device of the UHV converter station should operate, but the device itself does not operate. In the past, when there was no monitoring method, the on-site operators were unaware of this situation. After the monitoring facilities were installed, due to subjective factors such as the operator's experience and sense of responsibility, there would be cases of missed detection. Therefore, video analysis technology is now used to assist operators in determining the operation of the protection device.

[0003] The status signal parameters applicable to the core link of UHV DC protection that need to be monitored by the UHV converter station protection device are as follows: A. Export pressure plate status monitoring; B. Terminal block temperature measurement in the panel cabinet; C. Front panel monitoring of secondary equipment in the panel cabinet; D. Working temperature of secondary equipment in the panel cabinet; E. Working voltage of secondary equipment in the panel cabinet; F. Optical fiber light intensity monitoring; G. Cable insulation detection; H. Export loop detection; I. Auxiliary contact position; J. Cable status detection; K. Environmental parameter detection, such as temperature, humidity, etc.; L. Corrosion status of terminal blocks. The panel cabinet and outdoor terminal box are both UHV converter station protection devices. The most core protection information is fed back to the indoor panel cabinet and outdoor terminal box. Therefore, special attention should be paid to the protection status of the panel cabinet and outdoor terminal box. Among the above-mentioned UHV DC protection core link status signal parameters, the export pressure plate status monitoring and the front panel monitoring of secondary equipment in the panel cabinet are both in the panel cabinet, so image monitoring of the panel cabinet is required.

[0004] Video monitoring is mainly used for image monitoring of the cabinet, and video monitoring is widely used in UHV DC converter stations to monitor the operating status of equipment in each link. Because the information bandwidth of the image is limited, if all videos are transmitted to the cloud, the storage capacity and capacity of the cloud disk will be high. After the camera and lighting conditions are determined, the host computer has been collecting data under the same conditions. Therefore, the monitoring data in the small room of the UHV converter station is repetitive. The significance of uploading a large amount of repeated data is not great, and it occupies a large amount of storage space and transmission bandwidth in the cloud. Research on distributed data processing and fault analysis technology on the edge side of the UHV DC protection system based on lightweight artificial intelligence, which puts a large amount of data processed in the cloud on the edge side for processing, thereby achieving lightweight is an urgent problem to be solved in the current UHV converter station.

[0005] A glass cover is provided on the outside of the UHV converter station panel cabinet, and the entire panel cabinet is placed inside the glass cover. The panel cabinet is provided with a pressure plate, a test data display window, a switch, a handle, a status indicator light, etc. The status of the UHV converter station panel cabinet can be observed by taking photos of the UHV converter station panel cabinet through a camera installed in the small room of the UHV converter station. The information processing of monitoring images by traditional power equipment is mainly based on manpower, which is not efficient overall, and the accuracy of equipment operation status identification varies from person to person. The literature "Zhao Zhenbing, Zhang Wei, Zhai Yongjie, et al. Concept, Research Status and Prospect of Power Vision Technology [J]. Electric Power Science and Engineering, 2020, 36(01):1-8" proposed the concept of power vision technology, and established a bridge between power systems, computer vision and artificial intelligence. With the development of artificial intelligence and computer vision technology, a large number of UHV DC converter station equipment panoramic monitoring research works based on machine vision have emerged, which can achieve efficient and accurate acquisition of equipment operation status feature information and identification, and then complete effective video inspection work, saving a lot of manpower and improving inspection efficiency.

[0006] However, in the current video inspection process of converter station equipment, some specific scenes will show large-area reflection in the collected images due to the influence of reflected light, thus affecting the effect of image information processing. Although some scenes can be solved by changing the camera orientation, in the practice of secondary equipment cabinet monitoring, there is always the problem that the reflection of the cabinet surface glass makes it difficult to monitor some areas of the collected image (such as Figure 1 This greatly affects subsequent machine vision tasks such as target detection and semantic segmentation. In severe cases, the reflection interference area may even cover the entire area to be detected, making the task impossible to complete or interrupted. This is also an unresolved problem in the research on panoramic monitoring of UHV DC converter station equipment. Summary of the invention

[0007] The technical problem to be solved by the present invention is that in the panoramic monitoring process of UHV DC converter station equipment, there is a problem that the reflection of the surface glass of the panel cabinet makes it difficult to monitor part of the area of ​​the collected image.

[0008] The present invention solves the above-mentioned technical problems through the following technical means: a UHV converter station panel pressure plate monitoring and de-reflection interference network, comprising a plurality of sequentially cascaded encoders and a plurality of sequentially cascaded decoders, the encoder comprising two sequentially connected residual blocks and a WRNL block, the WRNL block output result of the current encoder is input into the residual block of the next-level encoder after downsampling, the decoder comprises sequentially connected convolutional layers, two residual blocks and a WRNL block, the convolutional layer is connected to an SE block, the output result of the previous level fused at the input end of the current decoder and the output result of each encoder are input into the SE block in the current decoder, the SE block adjusts the number of channels through the convolutional layer and then inputs the result into the residual block in the current decoder, the image with reflection interference is input into the first-level encoder, and the last-level decoder outputs the image without reflection interference.

[0009] The present invention uses a codec structure similar to a U-Net network as an image conversion model, realizes jump connections within the same level, and takes into account the element connections between levels. The output result of the previous level fused at the input end of the current decoder and the output result of each encoder are input into the SE block in the current decoder. The SE block adjusts the number of channels through a convolutional layer and then inputs the result into the residual block in the current decoder, thereby realizing interaction between different scales and avoiding information loss. The picture with reflection interference on the surface of the cabinet is input into the model, and the picture without reflection interference is output, thereby solving the problem that part of the collected image is difficult to monitor due to the reflection of the glass on the surface of the cabinet in the panoramic monitoring process of the UHV DC converter station equipment.

[0010] Furthermore, the de-reflection interference network includes three encoders cascaded in sequence and four decoders cascaded in sequence.

[0011] Furthermore, the residual block includes sequentially numbered first to third residual structures, the first to third residual structures are cascaded in sequence, each residual structure includes a 3×3 convolutional layer and a PReLU layer, the input end of the second residual structure receives the input data of the first residual structure and its output result, the input end of the third residual structure receives the input data of the second residual structure and its output result as well as the input data of the first residual structure, and the output data of the third residual structure fuses the input data of the first residual structure as the output of the entire residual block.

[0012] Furthermore, the working process of the WRNL block is:

[0013] The input feature X of WRNL is divided into a×b feature maps {X k},(k=1,...,ab), where k is the number of feature maps, through the formula Generate output features, where

[0014]

[0015]

[0016]

[0017]

[0018] Denote the output feature of the k-th feature map at position i. Denote the feature point of the i-th row in the k-th feature map. Denote the feature point of the j-th column in the k-th feature map, (). T Denote the transpose matrix of the matrix, W. θ and and W g are all weight matrices with dimensions C×L, C×L, and C×C respectively, and L = C / 2. Denote the correlation between i each in a set of regional positions S; γ(·) represents a relationship function and γ(·) = 1 / ((·)+1).

[0019] Furthermore, if a > b, the feature map is wider than when a = b. Therefore, when a > b, a = b, and a < b, they are respectively called a wide-region rectangular block, a square block, and a high-region rectangular block.

[0020] Further, after fusing the output result of the previous level and the output results of each encoder at the input end of the current decoder, the results are input into the SE block in the current decoder. After adjusting the number of channels through a convolutional layer, the results are input into the residual block in the current decoder, including:

[0021] Let be the output feature of level l (l = 1, 2, 3) in the encoder, then the input feature of each level l (l = 4, 5, 6, 7) in the decoder

[0022] where

[0023] represents the concatenation operation, H up (·) represents the upsampling operation, represents the decoder output feature of level l, W 1×1 represents a 1×1 convolutional layer, f SE (·) represents the SE block, represents the sampling operation from level i to l.

[0024] The present invention also provides a method for monitoring a de-reflection interference network of a UHV converter station panel pressure plate, the method comprising:

[0025] Build a dataset;

[0026] The de-reflection interference network is trained using the data set. When the convergence condition is met, a trained de-reflection interference network is obtained. The image with reflection interference collected in real time is input into the trained de-reflection interference network, and the image without reflection interference is output.

[0027] Furthermore, the data set includes a public data set and a panel pressure plate status image data set in a protection room of a UHV converter station, and the data set is randomly divided into a training set and a test set in a ratio of 7:3.

[0028] Furthermore, the convergence condition is that the loss function reaches a minimum value, and the loss function includes:

[0029] L 1 =||x gt -f(x input )|| 1 +||x gt -f(x input )|| 2

[0030] Among them, x input represents the input reflection image, x gt represents the corresponding de-reflected image, f represents the output of the de-reflected interference network, || || F Indicates F-norm calculation and F takes the values ​​of 1 or 2.

[0031] Furthermore, the loss function also includes:

[0032]

[0033] in,

[0034] θ represents the network weight, D represents the data set, I represents the input image, n is the image downsampling factor, Indicates that the transport layer pass factor is 2 n -1 bilinear interpolation downsampling, Indicates that the reflection layer has a pass factor of 2 n -1 bilinear difference downsampling, T represents the transmission layer of image I, R represents the reflection layer of image I, λ T and λ R are normalization factors, ⊙ means multiplication in unit order, represents the gradient map of the image I transmission layer, represents the gradient map of the reflection layer of image I, express The model, express Model;

[0035] The total loss function is L = L 1 +L excl (θ).

[0036] The advantages of the present invention are:

[0037] (1) The present invention uses a codec structure similar to a U-Net network as an image conversion model, realizes jump connections within the same level, and takes into account the element connections between levels. The output result of the previous level and the output result of each encoder are fused at the input end of the current decoder and then input into the SE block in the current decoder. The SE block adjusts the number of channels through a convolutional layer and then inputs the result into the residual block in the current decoder to achieve interaction between different scales and avoid information loss. The picture with reflection interference on the surface of the cabinet is input into the model, and the picture without reflection interference is output, which solves the problem that the reflection of the cabinet surface glass in the panoramic monitoring process of the UHV DC converter station equipment makes it difficult to monitor some areas of the collected image.

[0038] (2) Image dereflection is a low-level visual task that requires features with richer scales to restore details in the image. The present invention constructs a multi-level connection structure, where each level consists of two tightly connected residual blocks and a WRNL block. Each residual block consists of a first residual structure to a third residual structure. In the upsampling part of the network, feature information from all scales in the downsampling can be aggregated through multi-level connections. Since features at different levels have different scales, in order to adaptively adjust channel characteristics after multi-level connections, an SE block is added in each decoder stage, and the number of channels after the excitation block is adjusted through a 1×1 convolutional layer to match it with the input of the next level.

[0039] (3) The lighting characteristics in the small room of the converter station make conventional blocks inapplicable. The WENL block of the present invention divides the feature map into a grid with a wide area to make the reflection distribution relatively uniform. If the difference between the pixels in the input reflection image and the corresponding de-reflection image exceeds a certain threshold, the pixel is considered to belong to the reflection layer. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A schematic diagram of the structure of a UHV converter station panel pressure plate monitoring and de-reflection interference network provided by an embodiment of the present invention;

[0041] Figure 2A schematic diagram of the structure of a residual block in a UHV converter station panel pressure plate monitoring and de-reflection interference network provided by an embodiment of the present invention;

[0042] Figure 3 The multi-level connection level ablation experiment results of a public data set (PSNR1, SSIM1) and a data set of cabinet pressure plate status images (PSNR2, SSIM2) in a UHV converter station cabinet pressure plate monitoring and de-reflection interference network provided by an embodiment of the present invention;

[0043] Figure 4 A de-reflection visual processing result of a real natural landscape image in a UHV converter station panel pressure plate monitoring de-reflection interference network provided by an embodiment of the present invention;

[0044] Figure 5 The embodiment of the present invention provides a de-reflection visual processing result of the converter station cabinet pressure plate image in a UHV converter station cabinet pressure plate monitoring and de-reflection interference network. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] The reflected light source for monitoring the UHV DC converter station cabinet is mainly fluorescent light, which has the characteristics of indoor, non-natural light, scattered light spots and high light intensity. Figure 1 When inputting an image, the scattered light spots are likely to cover the display screen and the pressure plate waiting to be monitored. In addition, the installation position of the surveillance camera in the small room is restricted by environmental factors such as waterproof measures and cable routing. It can only be monitored by installing it on the top of the screen cabinet from a bird's-eye view angle. This makes it impossible to correctly read the device status information due to excessive light intensity, thereby affecting the image information processing effect and causing great interference to subsequent machine vision tasks such as target detection and semantic segmentation.

[0047] Existing de-reflection deep neural networks are mostly designed for real-world natural scenes, usually outdoors under natural light, where the intensity of reflected light is low and the light spot is evenly distributed, which is quite different from the lighting characteristics of the UHV DC converter station cabinets. Therefore, when constructing a de-reflection deep neural network for the UHV DC converter station cabinets, the operating environment and light characteristics of the equipment in the station must be fully considered to efficiently and accurately remove reflection interference and obtain characteristic information on the equipment operating status.

[0048] In view of the above-mentioned requirement for de-reflection of monitoring images, the present invention proposes the following Figure 1 The figure shows a UHV converter station panel pressure plate monitoring and de-reflection interference network, hereinafter referred to as the MA-Net network. The MA-Net network consists of an encoder and a decoder, the first three stages constitute the encoder part, and the remaining four stages constitute the decoder part. The levels are divided according to the size of the feature map, and a block is defined as a stage. Through the connection between multiple levels, MA-Net can connect all the outputs of the encoder to all the inputs of the decoder, so that features of different sizes can be used simultaneously in the process of image restoration. Continue to refer to Figure 1 The MA-Net network includes three sequentially cascaded encoders and four sequentially cascaded decoders. The encoder includes two sequentially connected residual blocks and a WRNL block. The output result of the WRNL block of the current encoder is input into the residual block of the next encoder after downsampling. The decoder includes sequentially connected convolutional layers, two residual blocks and a WRNL block. The convolutional layer is connected to an SE block. The output result of the previous level fused at the input end of the current decoder and the output result of each encoder are input into the SE block in the current decoder. The SE block adjusts the number of channels through the convolutional layer and then inputs the result into the residual block in the current decoder. The image with reflection interference is input into the first encoder, and the last decoder outputs the image without reflection interference. The following introduces the multi-level connection mechanism, WRNL block, network training process and loss function involved in the training in the above network structure in sections, and finally verifies the effect of the present invention through simulation.

[0049] 1. Multi-level connection mechanism

[0050] In a network structure similar to U-Net, the structure of connecting features at the same level can alleviate the defect that low-level features in the decoder cannot utilize multi-scale information. However, image dereflection is a low-level visual task that requires features with richer scales to restore details in the image. The present invention constructs a multi-level connection structure, each level of which consists of two densely connected residual (DCR) blocks (such as Figure 2It consists of WRNL blocks as shown. The residual block includes the first to the third residual structures numbered in sequence, and the first to the third residual structures are cascaded in sequence. Each residual structure includes a 3×3 convolutional layer and a PReLU layer. The input end of the second residual structure receives the input data and its output result of the first residual structure. The input end of the third residual structure receives the input data and its output result of the second residual structure and the input data of the first residual structure. The output data of the third residual structure fuses the input data of the first residual structure as the output of the entire residual block. In the upsampling part of the network, feature information from all scales in the downsampling can be aggregated through multi-level connections. Since features at different levels have different scales, in order to adaptively adjust the channel characteristics after multi-level connections, a squeeze-and-excitation (SE) block is added at each decoder stage, and the number of channels after the excitation block is adjusted through a 1×1 convolutional layer. Among them, the SE block uses the squeeze-and-excitation block of the existing technology, and for the relevant description of the squeeze-and-excitation block, reference can be made to the literature "Detailed Explanation of SE Module" published on the CSDN blog.

[0051] Let be the output feature of level l (l = 1, 2, 3) in the encoder, then the input feature of each level l (l = 4, 5, 6, 7) in the decoder is:

[0052] Among them,

[0053] Among them, represents the concatenation operation, H up (·) represents the upsampling operation, represents the decoder output feature of level l, W 1×1 represents a 1×1 convolutional layer, f SE (·) represents the SE block, represents the sampling operation from level i to l, that is, the l - i times of downsampling and i - l times of upsampling operations when l > i, l = i, and l < i.

[0054] Multi-level connections can use high-level features when processing low-level features, which helps the network to utilize multiple scale representations when restoring large-scale objects, and vice versa. The present invention uses discrete wavelet transform for upsampling and downsampling operations to find the mapping relationship between feature shapes at different scales. In addition, considering the problem of information loss, the present invention selects two-dimensional Haar wavelet for sampling operations.

[0055] 2. WRNL block (Wide Region Non-Local block)

[0056] The lighting characteristics in the small room of the converter station make conventional blocks inapplicable. Therefore, the WRNL block is first defined, and then statistical knowledge is used to analyze the effectiveness of the WRNL block.

[0057] The input feature X of WRNL is divided into a feature map of a×b {X k}, (k = 1,..., ab), where k is the number of feature maps. The output feature is generated through the formula where,

[0058]

[0059]

[0060]

[0061]

[0062] represents the output feature of the k-th feature map at position i, represents the i-th row feature point in the k-th feature map, represents the j-th column feature point in the k-th feature map, () T represents the transpose matrix of the matrix, W θ 、 and W g are all weight matrices with dimensions C×L, C×L, and C×C respectively, and L = C / 2; represents and the correlation between each i in a set of regional positions S ; γ(·) represents the relationship function and γ(·) = 1 / ((·)+1). If a > b, the grid is wider than when a = b. Therefore, when a > b, a = b, and a < b, they are called wide-region rectangular blocks, square blocks, and high-region rectangular blocks respectively.

[0063] Assuming that non-local blocks restore specific pixels based on the information of other pixels in the patch, each patch needs to have sufficient background information. However, due to the uneven distribution of the reflection layer, it is difficult for regional non-local blocks to fully utilize the background information. Wide-area rectangular patches have richer background information than square and high-area rectangular patches. The image is divided into 16*4, 8*8 and 4*16 grids, respectively, to obtain wide-area rectangular, square and high-area rectangular blocks. If the difference between the pixel in the input reflection image and the corresponding de-reflected image exceeds a certain threshold, the pixel is considered to belong to the reflection layer. The results of the ablation experiments of different region types are shown in Table 1. Compared with square and high-area rectangular patches, wide-area rectangular patches have better peak signal to noise ratio (PSNR) and structural similarity (SSIM). At this time, the distribution of the reflection layer on all patches is uniform, which can better restore the image.

[0064] Table 1 Region type ablation experiment of non-local blocks

[0065]

[0066] 3. Network training process

[0067] Construct a data set; use the data set to train the de-reflection interference network, and when the convergence condition is met, obtain a trained de-reflection interference network, input the real-time collected image with reflection interference into the trained de-reflection interference network, and output the image without reflection interference. The data set includes a public data set and a panel pressure plate state image data set in a UHV converter station protection room, and the data set is randomly divided into a training set and a test set in a ratio of 7:3.

[0068] The convergence condition is that the loss function reaches a minimum value, and the loss function includes:

[0069] L 1 =||x gt -f(x input )|| 1 +||x gt -f(x input )|| 2

[0070] Among them, x input represents the input reflection image, x gt represents the corresponding de-reflected image, f represents the output of the de-reflected interference network, || || F Indicates F-norm calculation and F takes the values ​​of 1 or 2.

[0071] In addition, in order to better separate the reflection layer and the transmission layer, a repulsion loss based on the gradient domain is defined. By analyzing the relationship between the edges of the two layers, it can be found that the transmission layer and the reflection layer basically do not overlap at the edge. The edge in image I should be caused by the transmission layer or the reflection layer, not the combination of the two. Therefore, the correlation between the transmission layer and the reflection layer predicted in the gradient domain is minimized, and the repulsion loss is expressed as the product of the normalized gradient fields of the two layers at multiple spatial resolutions. The loss function is constructed as follows:

[0072]

[0073] in,

[0074] θ represents the network weight, D represents the data set, I represents the input image, n is the image downsampling factor, Indicates that the transport layer pass factor is 2 n -1 bilinear interpolation downsampling, Indicates that the reflection layer has a pass factor of 2 n -1 bilinear difference downsampling, T represents the transmission layer of image I, R represents the reflection layer of image I, λ T and λ R are normalization factors, ⊙ means multiplication in unit order, represents the gradient map of the image I transmission layer, represents the gradient map of the reflection layer of image I, express The model, express The modulus of , N is equal to 3;

[0075] The total loss function is L = L 1 +L excl (θ).

[0076] 4. Simulation verification

[0077] 4.1 Dataset Selection

[0078] In order to verify the feasibility and effectiveness of the method of the present invention, the dataset SIR 2The proposed method is verified on the image dataset of the cabinet pressure plate status in the protection room of a UHV converter station. The proposed method uses the Pressure-plate (1400), Object (1500), Postcard (560) and Zhang et al. (800) datasets, totaling 4260 images, for training, and the remaining Pressure-plate (600), Object (640), Postcard (240) and Zhang et al. (340) datasets, totaling 1820 images, for quantitative evaluation. The four datasets are randomly divided into training set and test set in a ratio of 7:3.

[0079] The Pressure-plate dataset is an image dataset of the pressure plate state of the cabinet in an indoor environment, taken with a Canon EOS 750D camera. It contains 220 real image pairs, i.e., images with reflection and corresponding reference transmission layers. In order to simulate different imaging conditions, the following factors were considered when taking images: 1) Environment: indoor; 2) Lighting conditions: incandescent lamp; 3) Glass plate thickness: 3mm and 8mm; 4) Distance between glass and camera: 3-15cm; 5) Camera viewing angle: front and oblique; 6) Camera exposure value: 8.0-16.0; 7) Camera aperture (affects reflection blur): f / 4.0-f / 16.

[0080] 4.2 Experimental Results and Analysis

[0081] 4.2.1 Ablation Experiment

[0082] Through multi-level connections, feature information from all scales in the downsampling part can be aggregated in the upsampling part of the network. However, if the number of levels is too deep, the weight of key information will be reduced, and if the number of levels is too small, the effect of feature information extraction will not be obvious. Therefore, it is very important to choose the appropriate number of multi-level connections. Figure 3 The experimental results of multi-level connection series ablation on public datasets (PSNR1, SSIM1) and cabinet pressure plate state image datasets (PSNR2, SSIM2) are given. Figure 3 It can be seen that as the number of multi-level connections increases, the PSNR and SSIM indicators gradually increase, reaching the maximum value at level 4, and as the number of levels continues to increase, the PSNR and SSIM gradually decrease, indicating that the ability of the constructed deep neural network to aggregate information at various scales gradually decreases. Therefore, the number of subsequent multi-level connections is selected to be 4 layers.

[0083] 4.2.2 Qualitative analysis

[0084] Under the condition of selecting 4-layer connection level, the method MA-Net proposed in the present invention is compared with other methods, including CEILNet proposed in the document "Qingnan Fan, Jiaolong Yang, Gang Hua, et al. A generic deep architecture for single image reflection removal and image smoothing. IEEE International Conference on Computer Vision, 3238-3247, 2017.", BDN proposed in the document "Jie Yang, Dong Gong, Lingqiao Liu, et al. Seeing deeply and bidirectionally: A deep learning approach for single image reflection removal. European Conference on Computer Vision, 654-669, 2018." and ERRNet proposed in the document "Kaixuan Wei, Jiaolong Yang, Ying Fu, et al. Single image reflection removal exploiting misaligned training data and network enhancements. IEEE Conference on Computer Vision and Pattern Recognition, 8178-8187, 2019.". In order to make a peer comparison, the same public dataset training samples and the screen cabinet pressure plate state image dataset training samples are used to fine-tune the parameters of each model, and the best results of the fine-tuned version (indicated by the suffix '-F') are given.

[0085] like Figure 4 and Figure 5 As shown, the present invention shows the de-reflection visual processing results of real natural landscape images and converter station panel images, including input image (column 1), CEILNET (column 2), BDN (column 3), ERRNet (column 4), and the method of the present invention (column 5). It can be found that compared with other methods, the visual effect of the method of the present invention is more accurate, most of the unnecessary reflections are deleted, and the processing effect of indoor, non-natural light, scattered light spots and high light intensity is obviously superior, while other methods generally have problems such as unclear reflection removal effect and greater noise.

[0086] 4.2.3 Quantitative analysis

[0087] Table 2 summarizes the experimental results of different methods on four real datasets, including Pressure-plate, Object, Postcard, and Zhang et al. The number of test images in each dataset is shown after the name, and the PSNR and SSIM metrics are used. The larger the PSNR and SSIM values, the better the performance.

[0088] Table 2 Quantitative comparison of different methods on four real datasets

[0089]

[0090] As can be seen from Table 2, except for the Zhang et al. dataset published in the document "Xuaner Zhang, Ren Ng, Qifeng Chen. Single image reflection separation with perceptual losses. IEEE Conference on Computer Vision and Pattern Recognition, 4786-4794, 2018.", the MA-Net of the present invention achieves the best performance on all datasets. This is because ERRNet is based on the Zhang et al. model, and its network model has better generalization ability for the dataset, so the algorithm has better performance on the Zhang et al. dataset. In terms of the average performance of all test datasets, MA-Net is superior to other methods.

[0091] Table 3 compares and analyzes the influence of the de-reflection network on the pressure plate state recognition results based on the existing pressure plate state recognition methods (cluster matching method, improved BOF method, OpenVINO method, transfer learning method and improved SSD method) and the panel cabinet pressure plate state dataset Pressure-plate.

[0092] Table 3 Effect of de-reflection network on platen status recognition

[0093]

[0094] It can be seen from Table 3 that in the presence of reflection, the recognition accuracy of the five pressure plate state recognition methods are 78.22%, 83.55%, 92.90%, 89.63% and 84.55% respectively. After the de-reflection network removes the reflection interference, the recognition accuracy of the five methods has been improved to varying degrees. Among them, the recognition accuracy of the cluster matching method and the improved BOF method increased by 6.28% and 3.66% respectively, which is significantly higher than the 0.45%, 1.57% and 1.87% of the OpenVINO method, the transfer learning method and the improved SSD method. This is because the traditional image recognition method is more dependent on the information of the original image and has relatively poor anti-interference ability, so the de-reflection network produces better results.

[0095] 4.2.4 Experimental Conclusion

[0096] The present invention studies the reflection problem in the state monitoring of the pressure plate of the UHV DC converter station, and proposes a de-reflection deep neural network based on multi-level connection and adaptive regional attention to remove the reflection interference in the image. The MA-Net network can adaptively aggregate features through multi-level connection and compressed excitation blocks, and make full use of the rich remote non-reflection background information based on wide-region non-local blocks. Experiments show that MA-Net can not only restore the details of the input image, but also almost completely eliminate the reflection interference on the real image dataset and the panel cabinet pressure plate state image dataset, which can effectively improve the detection effect of the pressure plate state.

[0097] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A UHV converter station panel pressure plate monitoring and de-reflection interference network, characterized in that: The invention comprises a plurality of sequentially cascaded encoders and a plurality of sequentially cascaded decoders, wherein the encoder comprises two sequentially connected residual blocks and a WRNL block, the output result of the WRNL block of the current encoder is input into the residual block of the next encoder through downsampling, the decoder comprises sequentially connected convolutional layers, two residual blocks and a WRNL block, the convolutional layer is connected to an SE block, the output result of the previous level fused at the input end of the current decoder and the output result of each encoder are input into the SE block in the current decoder, the SE block adjusts the number of channels through the convolutional layer and then inputs the result into the residual block in the current decoder, the image with reflection interference is input into the first encoder, and the last decoder outputs the image without reflection interference; the working process of the WRNL block is as follows: The input feature X of WRNL is divided into a×b feature maps {X k },(k=1,...,ab), where k is the number of feature maps, through the formula Generate output features, where represents the output feature of the kth feature map at position i, represents the feature point in the i-th row of the k-th feature map, represents the jth column feature point in the kth feature map, () T represents the transposed matrix of the matrix, W θ , and W g They are all weight matrices with dimensions of C×L, C×L and C×C, respectively, with L=C / 2; express With a set of regional positions S i Each The correlation between them; γ(·) represents the relationship function and γ(·)=1 / ((·)+1).

2. According to claim 1, a UHV converter station panel pressure plate monitoring and de-reflection interference network is characterized in that: The de-reflection interference network includes three encoders cascaded in sequence and four decoders cascaded in sequence.

3. According to claim 1, a UHV converter station panel pressure plate monitoring and de-reflection interference network is characterized in that: The residual block includes sequentially numbered first to third residual structures, the first to third residual structures are cascaded in sequence, each residual structure includes a 3×3 convolutional layer and a PReLU layer, the input end of the second residual structure receives the input data of the first residual structure and its output result, the input end of the third residual structure receives the input data of the second residual structure and its output result as well as the input data of the first residual structure, and the output data of the third residual structure is fused with the input data of the first residual structure as the output of the entire residual block.

4. According to claim 1, a UHV converter station panel pressure plate monitoring and de-reflection interference network is characterized in that: If a>b, the feature map is wider than when a=b. Therefore, when a>b, a=b and a<b, they are called wide area rectangular blocks, square blocks and high area rectangular blocks respectively.

5. According to claim 1, a UHV converter station panel pressure plate monitoring and de-reflection interference network is characterized in that: The output result of the previous stage fused at the input end of the current decoder and the output result of each encoder are input into the SE block in the current decoder. The SE block adjusts the number of channels through the convolution layer and then inputs the result into the residual block in the current decoder, including: set up As the output feature of level l (l = 1, 2, 3) in the encoder, the input feature of each level l (l = 4, 5, 6, 7) in the decoder is for: in, Indicates cascade operation, H up (·) represents upsampling operation, represents the decoder output features of level l, W 1×1 represents a 1×1 convolutional layer, f SE (·) indicates SE block, Represents a sampling operation from level i to l.

6. A method for monitoring a de-reflection interference network of a UHV converter station panel pressure plate according to any one of claims 1 to 5, characterized in that: The method comprises: Build a dataset; The de-reflection interference network is trained using the data set. When the convergence condition is met, a trained de-reflection interference network is obtained. The image with reflection interference collected in real time is input into the trained de-reflection interference network, and the image without reflection interference is output.

7. A method for monitoring a de-reflection interference network of a UHV converter station panel pressure plate according to claim 6, characterized in that: The dataset includes a public dataset and a panel pressure plate status image dataset in a UHV converter station protection room. The dataset is randomly divided into a training set and a test set in a ratio of 7:

3.

8. The method for monitoring the de-reflection interference network of the UHV converter station panel pressure plate according to claim 6 is characterized in that: The convergence condition is that the loss function reaches a minimum value, and the loss function includes: L1=||x gt -f(x input )||1+||x gt -f(x input )||2 Among them, x input represents the input reflection image, x gt represents the corresponding de-reflected image, f represents the output of the de-reflected interference network, |||| F Indicates F-norm calculation and F takes the values ​​of 1 or 2.

9. A method for monitoring a de-reflection interference network of a UHV converter station panel pressure plate according to claim 8, characterized in that: The loss function also includes: in, θ represents the network weight, D represents the data set, I represents the input image, n is the image downsampling factor, Indicates that the transport layer pass factor is 2 n -1 bilinear interpolation downsampling, Indicates that the reflection layer has a pass factor of 2 n -1 bilinear difference downsampling, T represents the transmission layer of image I, R represents the reflection layer of image I, λ T and λ R are normalization factors, ⊙ means multiplication in unit order, represents the gradient map of the image I transmission layer, represents the gradient map of the reflection layer of image I, express The model, express Model; The total loss function is L = L1 + L excl (θ).

Images