Infrared image detection of insulators based on Centernet model

The infrared image can reflect the temperature information inside the insulator, which is helpful to realize the follow-up fault diagnosis. However, insulators are usually located in remote areas, and the complex background will affect the accuracy of the algorithm model. Therefore, we established an infrared dataset of insulators. We introduced the Centernet network to infrared detection of insulators, and conducted training and testing on the data set. The results show that our proposed model has a significant improvement of 4% to 5% compared with several mainstream detection models. This can meet the needs of insulator fault detection.


Loss Function
The overall loss function of the Centernet network is shown in formula (1), where is 0.1 and is 0.9. (1)

offset loss
Because the center point of the Centernet network has a certain floating point error in the prediction process. Therefore, the Centernet network uses the L1 Loss function to train the offset value of the center point. The specific formula is shown in (2).Among them, is the predicted bias value, and is the actual bias value of the network in the prediction process. The formula (2) reduces the floating point offset loss of the center point by comparing the L1 norm of the two, so it can realize more accurate prediction.

Center point prediction loss
The specific formula of the center point prediction loss function is shown in (3). Among them, and are hyperparameters, which are set to 2 and 4 in the Centernet network. N is the number of key points of the image. (3)

Width and height prediction loss
The width and height data of the target k and the corresponding center point coordinates are expanded according to formula (4). The Centernet network uses the value of to predict the center point, and at the same time realizes the regression of the width and height dimensions of each target. Finally, the L1 function is used as the loss function of the width and height prediction. Equation (5) is the width and height loss function, where is the predicted value, and is the length and width value after downsampling.

Experiment procedure
In order to ensure the reliability of the experimental data, the experiments in this paper are all trained and tested on the laboratory server. The server configuration is shown in Table 1. We run the Centernet network model in the Pytorch1.2 environment, training 32 samples in each batch, and training a total of 140 batches. The learning rate was initially set to 1.25×10 -4 , and was reduced to 0.1 times in the 90th and 120th batches respectively.

Result analysis
The speed detection standard of the deep learning detection model is the time it takes to detect a single image, in ms. Accuracy is measured by AP (average precision) value. The specific calculation formula is as formula (6).Among them, represents the ratio of the insulators successfully detected by the model to the total detection targets.
represents the completeness of the model′s feature detection of insulators. (true positives)represents the number of insulators correctly detected by the model. (false positives) represents the number of insulators identified as background. (false negatives) represents the number of insulators recognized as background. Finally, when analyzing the results, we will calculate the detection accuracy of the insulators and use the AP value to evaluate the overall performance of the model.

(6)
In the original Centeret, the author provided a total of four backbone networks: DLA34, Res18, Res101 and Hourglass. We compare these four types of backbone networks, and the results are shown in Table 2. It can be seen from the data in the table that Res18 has the highest accuracy among the four backbone networks, which has an accuracy advantage of 2%-9% compared to the other three networks. At the same time, the speed of Res18 is the highest among the four backbone networks, which is very We compared the Centernet model with Res18 as the backbone network with several mainstream detection models. As shown in Table 3, our network has a 4% to 5% improvement in accuracy compared to the three mainstream models, and at the same time performs best in terms of speed, slightly better than Yolov3. Figure 3 shows the Center point prediction map and the detection effect of our model on some datasets. On the whole, the overall performance of the Centernet model we proposed is due to several other mainstream models, which is very suitable for real-time fault diagnosis systems for power equipment.

Conclusion
The complex environment will affect the model's recognition rate of insulators. We introduce the Centernet network into insulator detection, and conduct training and testing on our own infrared data set. The results show that our proposed model has a significant improvement in accuracy of 4% to 5% compared with several mainstream models, and at the same time performs best in terms of speed. This will facilitate the real-time detection of infrared images of insulators in the next step.