Insulator identification system and method under backlight condition
By combining multi-scale fusion and interactive decoding of RGB and infrared image features, the problem of insulator identification under backlight conditions was solved, and high-precision insulator identification was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NORTH CHINA ELECTRIC POWER UNIV
- Filing Date
- 2022-07-26
- Publication Date
- 2026-05-15
AI Technical Summary
Under backlight conditions, it is difficult to obtain texture information from the RGB image captured by the visible camera for insulator identification, which makes it difficult to analyze the transmission line scene and identify the insulator.
The first and second unimodal feature pyramid decoders are used to extract features from RGB and infrared images, respectively. Information is fused through a feature fusion module, and semantic segmentation is performed using an interactive enhancement decoder. Multi-scale feature extraction and fusion are performed by combining the ResNet-50 backbone network and the feature pyramid network.
It effectively identifies insulators under backlight conditions, improving recognition accuracy and robustness, and reducing misidentification caused by the loss of texture features in backlight scenes.
Smart Images
Figure CN115223155B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image recognition technology, and in particular to an insulator identification system and method under backlight conditions. Background Technology
[0002] Insulators are critical components in high-voltage overhead transmission lines, supporting conductors and preventing current from flowing into the ground. However, insulators are prone to failure during prolonged operation in harsh outdoor conditions, such as rain, sunlight, and chemical corrosion. Therefore, timely inspection of insulator condition is crucial for transmission safety. With the development of artificial intelligence, intelligent insulator condition detection has become possible, replacing manual inspection. Identifying insulators from complex images is a prerequisite for condition inspection. Despite significant progress in recent years, the limitations of visible camera imaging, particularly the lack of texture information in RGB images obtained under backlight conditions, make transmission line scene analysis challenging, and insulator identification difficult under such conditions. Summary of the Invention
[0003] In view of this, the present invention provides an insulator identification system and method under backlight conditions to achieve insulator identification under backlight conditions.
[0004] To achieve the above objectives, the present invention provides the following solution:
[0005] An insulator identification system under backlight conditions, the identification system comprising: a first single-mode feature pyramid decoder, a second single-mode feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder; the first single-mode feature pyramid decoder and the second single-mode feature pyramid decoder are both connected to the feature fusion module; the feature fusion module is connected to the interactive enhancement decoder;
[0006] The first single-mode feature pyramid decoder is used to extract features from the RGB image of the insulator to obtain RGB image features;
[0007] The second single-mode feature pyramid decoder is used to extract features from the infrared image of the insulator to obtain infrared image features; both the RGB image and the infrared image of the insulator were obtained by shooting under backlight conditions;
[0008] The feature fusion module is used to fuse the RGB image features and the infrared image features to obtain the fused features;
[0009] The interactive enhancement decoder is used to decode and detect the fused features to obtain the semantic segmentation result of the insulator.
[0010] Optionally, the first unimodal feature pyramid decoder includes a ResNet-50 backbone network and a feature pyramid network connected in sequence.
[0011] Optionally, the second unimodal feature pyramid decoder includes an improved ResNet-50 backbone network and a feature pyramid network connected in sequence;
[0012] The improvement to the ResNet-50 backbone network is achieved by setting the number of input channels of the convolutional layers used to input the infrared image of the insulator in the ResNet-50 backbone network to 1, and deleting the average pooling layer and the fully connected layer in the ResNet-50 backbone network.
[0013] Optionally, the feature fusion module includes a multimodal feature fusion module and a global feature fusion module.
[0014] Optionally, the multimodal feature fusion module includes: a first concatenation module, a first convolution module, an X-dimensional average pooling layer, a Y-dimensional average pooling layer, a second concatenation module, a first convolutional layer, a BN and non-linear activation function layer, a second convolutional layer, a third convolutional layer, a first Sigmoid activation function layer, a second Sigmoid activation function layer, a multiplication operation module, a first channel attention module, a third concatenation module, and a fourth convolutional layer;
[0015] The first splicing module is connected to the first convolution module, and the first convolution module is connected to the X-dimensional average pooling layer, the Y-dimensional average pooling layer, and the first channel attention module, respectively.
[0016] Both the X-dimensional average pooling layer and the Y-dimensional average pooling layer are connected to the second splicing module. The second splicing module is connected to the first convolutional layer. The first convolutional layer is connected to the BN and non-linear activation function layers. The BN and non-linear activation function layers are connected to the second convolutional layer and the third convolutional layer, respectively. The second convolutional layer is connected to the first Sigmoid activation function layer. The third convolutional layer is connected to the second Sigmoid activation function layer.
[0017] The first Sigmoid activation function layer, the second Sigmoid activation function layer, and the first convolution module are all connected to the multiplication operation module;
[0018] The multiplication module and the first channel attention module are both connected to the third splicing module, and the third splicing module is connected to the fourth convolutional layer.
[0019] The first convolutional module includes a 1×1 convolution, a BN layer, and a ReLU activation function connected in sequence.
[0020] Optionally, the global feature fusion module includes a fourth concatenation module, a second channel attention module, a second convolution module, a global adaptive pooling layer, a third convolution module, an upsampling layer, a fourth concatenation module, and a fourth convolution module connected in sequence; the second convolution module is also connected to the fourth concatenation module;
[0021] Both the second convolutional module and the fourth convolutional module include a 1×1 convolution, a BN layer, and a ReLU activation function connected in sequence;
[0022] The third convolutional module includes multiple 1×1 convolutions, each of which is connected to a global adaptive pooling layer and an upsampling layer, respectively.
[0023] An insulator identification method under backlight conditions, the identification method being applied to the aforementioned identification system, the identification method comprising the following steps:
[0024] The RGB and infrared images of the insulators were acquired; both images were taken under backlighting conditions.
[0025] The RGB image features of the insulator are extracted by the first single-mode feature pyramid decoder.
[0026] The infrared image features of the insulator are extracted by using a second single-mode feature pyramid decoder.
[0027] The RGB image features and the infrared image features are fused by the feature fusion module to obtain the fused features;
[0028] The semantic segmentation results of the insulator are obtained by decoding and detecting the fused features through an interactive enhanced decoder.
[0029] Optionally, the step of extracting features from the RGB image of the insulator using a first single-mode feature pyramid decoder to obtain RGB image features specifically includes:
[0030] Multi-scale features of RGB images are extracted using the ResNet-50 backbone network in the first unimodal feature pyramid decoder;
[0031] The multi-scale features of the RGB image are fused by the feature pyramid network in the first unimodal feature pyramid decoder to obtain the multi-scale fused features of the RGB image as the RGB image features.
[0032] Optionally, the fused features include modal fusion features and global fusion features. The step of fusing the RGB image features and the infrared image features through a feature fusion module to obtain the fused features specifically includes:
[0033] The RGB image features and the infrared image features are fused using a multimodal feature fusion module to obtain modal fusion features;
[0034] The global feature fusion module performs global fusion of the RGB image features and the infrared image features to obtain global fused features.
[0035] Optionally, the step of extracting features from the RGB image of the insulator using a first single-mode feature pyramid decoder to obtain RGB image features further includes:
[0036] Acquire RGB and infrared images of insulators under backlighting conditions and construct a dataset;
[0037] The dataset is divided into a training set, a validation set, and a test set;
[0038] Using the training set, the validation set, and the test set, the network structure is trained, validated, and tested on the NVIDIA RTX 3090 platform to obtain the trained network structure as the recognition system. The network structure includes a first unimodal feature pyramid decoder, a second unimodal feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder.
[0039] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0040] This invention discloses an insulator identification system and method under backlight conditions. The identification system includes: a first single-mode feature pyramid decoder, a second single-mode feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder. Both the first and second single-mode feature pyramid decoders are connected to the feature fusion module, which is connected to the interactive enhancement decoder. This invention addresses the limitation of limited texture features in visible light imaging under backlight conditions by utilizing infrared images to supplement the information in the RGB images. Features are extracted from the RGB and infrared images of the insulator using the first and second single-mode feature pyramid decoders, respectively. The feature fusion module then fuses these features to enhance their information. Finally, the fused features are fed into the interactive decoder for image recovery and detection. This invention enables effective identification of insulators under backlight conditions. Attached Figure Description
[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0042] Figure 1 This is a network structure diagram of an insulator identification system under backlight conditions provided in an embodiment of the present invention;
[0043] Figure 2 This is a network structure diagram of the multimodal feature fusion module provided in an embodiment of the present invention;
[0044] Figure 3 This is a network structure diagram of the global feature fusion module provided in an embodiment of the present invention;
[0045] Figure 4 This is a network structure diagram of the interactive enhancement decoder provided in an embodiment of the present invention;
[0046] Figure 5 A comparison diagram of recognition effects provided in the embodiments of the present invention; Figure 5 In the figures (a)-(e), respectively: RGB image example, infrared image example, real annotation example, recognition result example of the existing method RTFNet, and recognition result example of the method of the present invention. Detailed Implementation
[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0048] The purpose of this invention is to provide an insulator identification system and method under backlight conditions, so as to realize the identification of insulators under backlight conditions.
[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0050] Example 1
[0051] like Figure 1As shown, Embodiment 1 of the present invention provides an insulator identification system under backlight conditions. The identification system includes: a first single-mode feature pyramid decoder, a second single-mode feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder. The first single-mode feature pyramid decoder and the second single-mode feature pyramid decoder are both connected to the feature fusion module. The feature fusion module is connected to the interactive enhancement decoder. The first single-mode feature pyramid decoder is used to extract features from the RGB image of the insulator to obtain RGB image features. The second single-mode feature pyramid decoder is used to extract features from the infrared image of the insulator to obtain infrared image features. Both the RGB image and the infrared image of the insulator are obtained by shooting under backlight conditions. The feature fusion module is used to fuse the RGB image features and the infrared image features to obtain fused features. The interactive enhancement decoder is used to decode and detect the fused features to obtain the semantic segmentation result of the insulator.
[0052] The purpose of the first and second unimodal feature pyramid decoders is to extract features from dual-light images (RGB and infrared images), enabling the network to learn the features of targets in the images. Both the first and second unimodal feature pyramid decoders include a ResNet-50 backbone network and a feature pyramid network.
[0053] (1) The two-light images are first passed through the ResNet-50 backbone network to extract multi-scale features of the RGB images. Multi-scale features of infrared images {T i 1 |i=1,2,3,4,5}. However, ResNet-50 is designed for three-channel RGB images, while infrared images have only one channel. Therefore, in the ResNet-50 network for infrared image feature extraction, the input channel size of the initial convolutional layer is modified to 1. Simultaneously, the average pooling layer and fully connected layer of ResNet-50 are removed to avoid the loss of spatial information, resulting in the improved ResNet-50 backbone network. Here, i represents the i-th scale feature.
[0054] (2) Because lower-level features have less semantic information but accurate target location, while higher-level features have richer semantic information but coarser target location, the multi-scale features of the RGB image after passing through the ResNet-50 backbone network are used to... Multi-scale features of infrared images {T i 1The features of |i=1,2,3,4,5} are further fused through a feature pyramid network to improve network performance. The fused features are: and {T i 2 |i=2,3,4,5}.
[0055] The purpose of the feature fusion module is to complement the features of RGB and infrared images to obtain robust or information-rich features. The feature fusion module mainly consists of two parts:
[0056] 1) Multimodal Feature Fusion Module (MFFM):
[0057] like Figure 2 As shown, firstly, the features of the RGB image are analyzed. and infrared image features {T i 2 The features |i=2,3,4} are concatenated along the channel dimension using the first concatenation module. Then, the features are integrated using a first convolutional module consisting of a 1×1 convolution, a BN layer, and a ReLU activation function to obtain feature u. j ; respectively, average pooling layers in the X and Y dimensions are used to average pool u in the X and Y dimensions. j After performing average pooling, the layers are concatenated along the channel dimension using a second concatenation module. Then, f is obtained through a 1×1 convolution (first convolutional layer), a batch normalization (BN) layer, and a non-linear activation function (BN and non-linear activation function layers). j Then f j The vector is divided into two parts in the horizontal and vertical dimensions, and then passed through 1×1 convolutions (second and third convolutional layers) and Sigmoid activation functions (first and second Sigmoid activation function layers) respectively. j Multiply to get At the same time, u j Through channel attention (first channel attention module), the more important features in the channel dimension are learned by means of feature compression and activation. Finally and The third stitching module performs stitching along the channel dimension, and the stitched features are then reduced in dimensionality using a 1×1 convolution (the fourth convolutional layer) to obtain the fused feature {F}. j |j=2,3,4}.
[0058] 2) Global Feature Fusion Module (GFFM):
[0059] like Figure 3 As shown, for RGB image features and infrared image features T52 The fourth concatenation module concatenates features along the channel dimension. Then, the second channel attention module learns the more important features along the channel dimension through feature compression and activation. Next, a second convolutional module, consisting of a 1×1 convolution, a BN layer, and a ReLU activation function, integrates the features to obtain feature map v. Feature v is then processed by a global adaptive pooling layer to obtain feature maps of 1×1, 5×5, 9×9, and 13×13, respectively. These are then passed through 1×1 convolutional layers and upsampled to restore the original feature map size, resulting in four feature maps {w}. n |n=1,2,3,4};Connect feature map v and feature map w n The n = 1, 2, 3, 4 are concatenated along the channel dimension, and then passed through a 1×1 convolution, a BN layer, and a ReLU activation function to obtain the fused feature F5.
[0060] like Figure 4 As shown, the interactive enhancement decoder gradually restores the resolution of the fused feature map F5 to the size of the original input image. During the feature decoding process, it interacts with the fused feature map {F... j The features are progressively added from |j=2,3,4} to compensate for the feature information lost during the upsampling process of F5. Once the feature map is restored to its original size, it is passed through a Softmax layer to obtain a semantic probability map for the segmentation result.
[0061] Example 2
[0062] Embodiment 2 of the present invention provides an insulator identification method under backlight conditions. The identification method is applied to the above-mentioned identification system and includes the following steps:
[0063] The RGB and infrared images of the insulators were acquired; both images were taken under backlighting conditions.
[0064] The RGB image of the insulator is extracted using a first single-modal feature pyramid decoder to obtain RGB image features. Specifically, the extraction of RGB image features using the first single-modal feature pyramid decoder includes: extracting multi-scale features of the RGB image using the ResNet-50 backbone network in the first single-modal feature pyramid decoder; and fusing the multi-scale features of the RGB image using the feature pyramid network in the first single-modal feature pyramid decoder to obtain multi-scale fused features of the RGB image as RGB image features.
[0065] The infrared image features of the insulator are extracted by using a second single-mode feature pyramid decoder.
[0066] The RGB image features and the infrared image features are fused using a feature fusion module to obtain fused features. Specifically, the fusion of the RGB image features and the infrared image features using a multimodal feature fusion module to perform modal fusion of the RGB image features and the infrared image features to obtain modal fused features, and the global feature fusion module to perform global fusion of the RGB image features and the infrared image features to obtain global fused features.
[0067] The semantic segmentation results of the insulator are obtained by decoding and detecting the fused features through an interactive enhanced decoder.
[0068] The step of extracting features from the RGB image of the insulator using a first single-modal feature pyramid decoder to obtain RGB image features includes, prior to: acquiring RGB and infrared images of the insulator in a backlit scene to construct a dataset; dividing the dataset into a training set, a validation set, and a test set; and using the training set, the validation set, and the test set, training, validating, and testing the network structure on an NVIDIA RTX 3090 platform to obtain the trained network structure as the recognition system. The network structure includes a first single-modal feature pyramid decoder, a second single-modal feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder.
[0069] The dataset is constructed as follows:
[0070] A paired visible light and infrared (RGB-T) insulator dataset with backlit scenes was used. The RGB-T insulator dataset for backlit scenes was acquired using a FLIR T600 camera, which can simultaneously transmit RGB and infrared images. The dataset contains 1426 pairs of RGB and thermal images, with resolutions of 360×480 and 480×640 for the RGB and infrared images, respectively. The dataset was randomly split according to a 50% training set, 25% validation set, and 25% test set split. The proposed model was trained and tested on an NVIDIA RTX 3090 platform running Ubuntu 18.04.1LTS. Training was accelerated using CUDA 11.2, with Python 3.7 as the programming language and PyTorch as the network framework. Pre-trained weights for ResNet were loaded using PyTorch. During the decoding phase, the transposed convolutional layers were initialized. Training was performed using stochastic gradient descent (SGD) with a momentum of 0.9 and weight decay of 0.0005. The initial learning rate was set to 0.01 and gradually decreased using an exponential decay scheme. Batch size was adjusted based on the specific network. Input images were resized to 480×640 pixels, and a flipping method was used to augment the dataset.
[0071] To evaluate the effectiveness of the proposed model, this invention employs two commonly used evaluation metrics: accuracy (Acc) and intersection (IoU). The calculation formulas are as follows:
[0072]
[0073]
[0074] In the formula, C and K represent the number of classes and the number of test images, respectively. In test image k, This represents the number of pixels that are correctly classified into class i. This represents the number of pixels that were misclassified from class i to class j. This represents the number of pixels that were misclassified from class j to class i. The average values for all classes are represented by mIoU and mAcc.
[0075] To illustrate the effectiveness of the identification system and method of the present invention, the present invention Figure 5 The image shows a comparison of the recognition results obtained using the method of this invention. Figure 5 In the figures (a)-(e), respectively: an example image of an RGB image, an example image of an infrared image, an example image of a ground truth annotation, an example image of the recognition result of the existing method RTFNet, and an example image of the recognition result of the method of the present invention. Figure 5In each group of images (a)-(e), the three images represent different scenes, with the first row showing mild backlighting, the second row showing moderate backlighting, and the third row showing severe backlighting. Figure 5 In comparison with other methods, the method of this invention demonstrates superior results in challenging backlit scenes. Particularly in the second and third rows, the texture features of insulators are lost in backlit scenes, making them easily confused with towers. However, the method of this invention demonstrates its robustness, avoiding misidentification of towers as insulators. Therefore, the system and method of this invention can effectively identify insulators under backlit conditions with excellent accuracy.
[0076] This invention addresses the limitations of visible light imaging by utilizing thermal information to supplement RGB information. First, the captured visible light and infrared images are fed into a single-modal feature pyramid encoder for feature extraction. Then, the extracted visible light and infrared image features are fed into a feature fusion module for feature fusion to enhance the feature information. Finally, the fused features are fed into an interactive enhancement decoder for image restoration and detection.
[0077] Experimental results demonstrate the superiority of this network in backlight scene analysis, with detection results outperforming other models. This paper proposes a multimodal feature fusion method, namely the fusion of visible light images and infrared images, providing a new approach for insulator detection under backlight conditions.
[0078] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0079] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. An insulator identification system under backlight conditions, characterized in that, The recognition system includes: a first unimodal feature pyramid decoder, a second unimodal feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder; the first unimodal feature pyramid decoder and the second unimodal feature pyramid decoder are both connected to the feature fusion module; the feature fusion module is connected to the interactive enhancement decoder; The first single-mode feature pyramid decoder is used to extract features from the RGB image of the insulator to obtain RGB image features; The second single-mode feature pyramid decoder is used to extract features from the infrared image of the insulator to obtain infrared image features; both the RGB image and the infrared image of the insulator were obtained by shooting under backlight conditions; The feature fusion module is used to fuse the RGB image features and the infrared image features to obtain the fused features; The interactive enhancement decoder is used to decode and detect the fused features to obtain the semantic segmentation result of the insulator; The feature fusion module includes a multimodal feature fusion module and a global feature fusion module; The multimodal feature fusion module includes: a first concatenation module, a first convolution module, an X-dimensional average pooling layer, a Y-dimensional average pooling layer, a second concatenation module, a first convolution layer, a BN and non-linear activation function layer, a second convolution layer, a third convolution layer, a first Sigmoid activation function layer, a second Sigmoid activation function layer, a multiplication operation module, a first channel attention module, a third concatenation module, and a fourth convolution layer; The first splicing module is connected to the first convolution module, and the first convolution module is connected to the X-dimensional average pooling layer, the Y-dimensional average pooling layer, and the first channel attention module, respectively. Both the X-dimensional average pooling layer and the Y-dimensional average pooling layer are connected to the second splicing module. The second splicing module is connected to the first convolutional layer. The first convolutional layer is connected to the BN and non-linear activation function layers. The BN and non-linear activation function layers are connected to the second convolutional layer and the third convolutional layer, respectively. The second convolutional layer is connected to the first Sigmoid activation function layer. The third convolutional layer is connected to the second Sigmoid activation function layer. The first Sigmoid activation function layer, the second Sigmoid activation function layer, and the first convolution module are all connected to the multiplication operation module; The multiplication module and the first channel attention module are both connected to the third splicing module, and the third splicing module is connected to the fourth convolutional layer. The first convolutional module includes a 1×1 convolution, a BN layer, and a ReLU activation function connected in sequence.
2. The insulator identification system under backlight conditions according to claim 1, characterized in that, The first unimodal feature pyramid decoder includes a ResNet-50 backbone network and a feature pyramid network connected in sequence.
3. The insulator identification system under backlight conditions according to claim 1, characterized in that, The second unimodal feature pyramid decoder includes an improved ResNet-50 backbone network and a feature pyramid network connected in sequence; The improvement to the ResNet-50 backbone network is achieved by setting the number of input channels of the convolutional layers used to input the infrared image of the insulator in the ResNet-50 backbone network to 1, and deleting the average pooling layer and the fully connected layer in the ResNet-50 backbone network.
4. The insulator identification system under backlight conditions according to claim 1, characterized in that, The global feature fusion module includes a fourth concatenation module, a second channel attention module, a second convolution module, a global adaptive pooling layer, a third convolution module, an upsampling layer, a fourth concatenation module, and a fourth convolution module connected in sequence; the second convolution module is also connected to the fourth concatenation module. Both the second convolutional module and the fourth convolutional module include a 1×1 convolution, a BN layer, and a ReLU activation function connected in sequence; The third convolutional module includes multiple 1×1 convolutions, each of which is connected to a global adaptive pooling layer and an upsampling layer, respectively.
5. A method for identifying insulators under backlight conditions, characterized in that, The identification method is applied to the identification system according to any one of claims 1-4, and the identification method includes the following steps: The RGB and infrared images of the insulators were acquired; both images were taken under backlighting conditions. The RGB image features of the insulator are extracted by the first single-mode feature pyramid decoder. The infrared image features of the insulator are extracted by the second single-mode feature pyramid decoder. The RGB image features and the infrared image features are fused by the feature fusion module to obtain the fused features; The semantic segmentation results of the insulator are obtained by decoding and detecting the fused features through an interactive enhanced decoder.
6. The insulator identification method under backlight conditions according to claim 5, characterized in that, The step of extracting features from the RGB image of the insulator using a first single-mode feature pyramid decoder to obtain RGB image features specifically includes: Multi-scale features of RGB images are extracted using the ResNet-50 backbone network in the first unimodal feature pyramid decoder; The multi-scale features of the RGB image are fused by the feature pyramid network in the first unimodal feature pyramid decoder to obtain the multi-scale fused features of the RGB image as the RGB image features.
7. The insulator identification method under backlight conditions according to claim 5, characterized in that, The fused features include modal fusion features and global fusion features. The process of fusing the RGB image features and the infrared image features through a feature fusion module to obtain the fused features specifically includes: The RGB image features and the infrared image features are fused using a multimodal feature fusion module to obtain modal fusion features; The global feature fusion module performs global fusion of the RGB image features and the infrared image features to obtain global fused features.
8. The insulator identification method under backlight conditions according to claim 5, characterized in that, The step of extracting features from the RGB image of the insulator using a first single-mode feature pyramid decoder to obtain RGB image features also includes the following prior steps: Acquire RGB and infrared images of insulators under backlighting conditions and construct a dataset; The dataset is divided into a training set, a validation set, and a test set; Using the training set, the validation set, and the test set, the network structure is trained, validated, and tested on the NVIDIA RTX 3090 platform to obtain the trained network structure as the recognition system. The network structure includes a first unimodal feature pyramid decoder, a second unimodal feature pyramid decoder, a feature fusion module, and an interactive enhancement decoder.