A Method for Image Recognition of Insulator Jacket Defects under Small Sample Conditions

Accurate identification of insulator jacket defect images requires a large number of samples for model training, and the actual defect image datasets available for model training is seriously insufficient. In order to solve the problems of the model cannot be trained, over-fitting and low accuracy caused by too few training samples, this paper proposes a new method for image recognition of insulator jacket defects under small sample conditions, which combines image enhancement technology and meta-learning technology to train the U-Net image segmentation network, and finally obtain the image recognition model of the insulator jacket defect. In this paper, the defect recognition models using meta-learning method and without meta-learning are compared experimentally, and the results show that the proposed method can achieve accurate recognition with a small-scale original data set.


Introduction
At present, the accurate recognition of insulator jacket defect images requires a large number of samples for model training, and the actual defect image dataset available for model training is seriously insufficient, which leads to the problems of poor defect fault recognition accuracy and training overfitting. In this paper, we propose a data processing method based on image enhancement transform as well as meta-learning, and combine it with U-Net image segmentation network to achieve the purpose of insulator defect recognition of power equipment under the condition of small samples. [1][2] This paper proposes a small sample insulator jacket defect image recognition method idea, as shown in Figure 1.
In this paper, the defect recognition models using meta-learning method and without meta-learning are experimentally compared respectively, and the results show that the method proposed in this paper can achieve accurate recognition of insulator jacket defect images with small-scale original data set. Defect recognition was performed on 180 insulator images, of which 36 had crack defects, and 30 of them could be recognized without using the meta-learning method, with a success rate of 83.3%; the present method could accurately recognize 36 of them, with a success rate of 100%.The proposed method consists of the following processing procedures: 1.expand the small sample insulator jacket image dataset using data enhancement methods such as image rotation, local magnification and Gaussian blurring; 2. extract the weights and model features of the pre-trained model on the source domain dataset (outside the fully connected layer) and train it on the target dataset using migration learning methods; 3. fine-tune the fine-tune the model trained on the target dataset to obtain the final insulator defect recognition model.

Insulator Jacket Defect Image Dataset Enhancement Method
In this paper, we use image rotation, local magnification method and Gaussian blur image processing technique to expand the defect image data set and solve the problem of insufficient amount of defect image data. Image rotation is the process of rotating a point on a graph as a fixed point by a certain angle to produce a new graph, and its coordinate calculation schematic and effect, Figure 2. [2][3] The effect of partially enlarging the number of pixels in some areas of an image to highlight the features of some areas is called partial image enlargement. Schematic diagram and effect of image local enlargement processing. The results are shown in Figure 2.
a-Rotation processing of defective insulator images b-Partial enlargement of insulator picture

U-Net Network for Insulator Jacket Defect Identification
U-Net network is a fully convolutional neural network. The advantages over ordinary fully convolutional neural networks are: 1. High correctness can be achieved with small data sets, suitable for small sample data set applications; 2. Simple network structure; 3. Through the process of downsampling and upsampling, the feature map is guaranteed to contain more detailed features and the image edges are more refined.
U-Net consists of two parts, downsampling and upsampling. The left half is a downsampling process, also known as feature extraction, which consists of two 3x3 ReLu convolutional layers and a 2x2 max pool pooling layer iteratively, forming a downsampling process. The right half is an upsampling process, also known as extended path, which consists of a deconvolutional layer, feature splicing and two 3x3 ReLu convolutional layers iteratively, forming an upsampling process, and finally outputting the feature segmentation results, ensuring that the finally recovered feature map incorporates more detailed features and also incorporates features of different scales, the U-Net network structure, as shown in Figure 4-a.
a-U-net network diagram b-insulator jacket image feature segmentation process  Take the insulator jacket image processing as an example, at first, the network inputs a 700×700 insulator jacket image, and after downsampling, it will get the feature maps of 696×696, 344×344, 168×168, 80×80 and 36×36 respectively. This 72×72 feature map is merged with the previous 80×80 feature map, and the 80×80 feature map is cropped to fit the size of the 72×72 feature map during the merging process, and then the spliced feature map is convolved and upsampled to obtain the 136×136 feature map. After splicing with the previous 168×168 feature map, convolution, upsampling and splicing, a total of four upsampling processes are performed, and finally a prediction result map of size 516×516 is obtained, as shown in Figure 4-b.

Training of Insulator Jacket Defect Recognition U-Net
Meta-learning is a neural network training method that enables models to "learn to learn". Its core idea is task-based learning, by learning features representation of a task to generalize on a new task. [4]

Meta-Learning Based Training Methods
The process of meta-learning is: the initial network parameters with strong generalization capability are obtained from the training and validation datasets, the network model will perform several gradient descent operations in the test task to achieve the purpose of learning a new task, and then the effect of the learning is verified, the training and testing process is schematically shown in Figure 5 By using a meta-learning method to train the U-Net neural network, a network with more potential and easier convergence to a global loss optimum is obtained, and then a new task is learned in the target application domain (insulator jacket defect dataset) to obtain a network suitable for insulator jacket defect image recognition. Solving the problems of insufficient generalization ability, overfitting and poor adaptability to new tasks in small sample data sets of traditional machine learning neural network.

Meta-Learning Based Training Process
In this paper, meta-learning training is performed on a small sample insulator jacket defect dataset, and the insulator jacket defect recognition network training process is shown in Figure 5-b.In Figure 5-b, the insulator jacket defect image recognition network training process includes the following key processes: (1) First, the insulator jacket defect image data set is expanded using image selection, image local magnification and Gaussian blurring methods. (2) Second, the insulator jacket defect image is  3) Third, a support set, a query set and a test set, are generated according to the meta-learning approach. (4) Fourth, loading the U-Net network trained from the concrete crack defect image dataset, at which point the network has the ability to learn the task, initializing its parameters and starting training on the insulator jacket defect dataset. (5) Fifth, after the training, the network model is tested and validated, and the pre-processed insulator images are input into the U-Net network for prediction, and the output Feature Map is obtained; finally, according to the output confidence level in the output Feature Map, combined with the threshold value set by the algorithm, the existence of crack features is judged; the feature information such as the location, length and width of the crack can be obtained from the Feature Map, and the final model is obtained after verification.

Method Validation and Analysis
In this paper, a workstation is chosen as the experimental platform. The processor is Intel Xeon E5, memory 128GB, 8 GPU graphics card GTX2080TI. The software framework structure is based on the deep learning framework of Tensor Flow 2.0. In the experiments of this paper, Kodak SL10 detachable miniature smart lens was selected for insulator jacket defect image acquisition, and for higher insulator devices, the lens was fixed to the insulator pole and the lens was controlled via Bluetooth. Image acquisition equipment diagram, as shown in Figure 6

Selection of Network Training Data
In this paper, the experimental source domain cracking experimental data is adopted from the concrete cracking defect image dataset published on the web, and the data is divided into support set, query set and test set according to the meta-learning method, and the division of the dataset is detailed in Table  1. Through the image acquisition equipment is used, the insulator jacket defects were photographed and collected, and the collected substation equipment insulator pictures were used as the original samples, and the original dataset was expanded according to the data enhancement method described in Subsection 1 to obtain the enhanced dataset of about 2000 pictures, and the dataset was labeled, and finally split into Support Set, Query Set, and Test Set. The before and after processing of the substation insulator dataset, the graphs are shown in Figure 6-b and Table 2.The training parameters of the U-Net model used in the training process are shown in Table 3.

Defect Recognition Accuracy and Model Convergence Comparison Analysis
The defect recognition models using meta-learning methods and those without meta-learning are experimentally compared respectively, and the U-Net is selected with a batch size of 6 and an iteration number of 300 for training. The accuracy metric curves and loss function curves of U-Net trained and tested by different learning methods are shown in Figure 7~Figure 8.
a-Not using meta-learning accuracy metric curves b-No meta-learning loss function curve Combining Figure 7~Figure 8, we can analyze that: 1. from the U-Net accuracy metric curve, using meta-learning method and not using meta-learning method do not have a great impact on U-Net network accuracy; 2. from the loss function curve, after using meta-learning method, the convergence speed of U-Net is obviously faster than not using meta-learning method. The results show that the recognition success rate is significantly lower without the meta-learning method than with the meta-learning method, as shown in Table 4.  The final model of insulator jacket defect recognition is obtained using the method of this paper, and the defect recognition experiment is carried out on 180 insulator pictures, and the results of crack defect recognition during the experiment are shown in Figure 9. Through the experiment, it can be obtained that using the meta-learning method to train the insulator jacket defect network can effectively shorten the time of network convergence, and at the same time the efficiency and accuracy of the network model are also improved. From Fig. 8, it can be obtained that the accuracy of the model for recognizing insulator jacket crack defects reaches 100%. In the experiment of 180 insulator pictures for defect recognition, 36 of them have crack defects in insulators, and this method can accurately recognize 36 of them, with a success rate of 100%.

Conclusion
In this paper, we propose to expand the dataset by augmented transformation based on a small dataset, and introduce a neural network pre-training model based on meta-learning to achieve the recognition of insulator jacket defects. The conclusions are obtained as follows:(1) The problem of insufficient training data for the neural network model is solved by expanding the size of the original small sample insulator jacket dataset through image rotation, image local magnification, and image enhancement transformations such as Gaussian filtering.(2) The use of meta-learning through neural network learning constructs a network with more potential and easier convergence to a global loss optimum. Solve the problem of insufficient generalization ability and overfitting of traditional neural network learning. (3) The proposed method also provides new ideas for fault diagnosis of transformer operating insulators, unmanned substations and intelligent recognition of cracked insulators by robots.