Power Detection Image Recognition Analysis Based on Image Fusion Extended Training

The operation and maintenance of power system are paid more and more attention. How to solve the contradiction between economic development and electric power construction is the key problem faced by the development of the current national power industry. Image recognition technology realizes the transformation from artificial to intelligent power grid in power system operation and maintenance. With the help of computer vision algorithms and deep learning modes, the automatic detection and remote control of power system equipment and lines are realized. Image recognition techniques generate desired image samples through image synthesis algorithms, but the lack of image samples in power inspection leads to poor performance of the recognition algorithm. To solve this problem, this chapter gives two methods to expand the sample of power inspection. Firstly, the image synthesis algorithm based on image fusion is introduced. For some samples, the Poisson fusion algorithm is used to smooth the gradient of the boundary. Then, the image generation algorithm is introduced and the required samples are directly generated through Deep Convolutional Generative Adversarial Networks (DCGAN). Finally, two sample set augmentation methods are evaluated by experiments. Experiments are conducted to evaluate the effect of the two sample augmentation methods. In the experiment, the F1-score of the trained detection model on the same test set is improved from 0.802 to 0.873 after using image fusion to expand the training set. After using the images generated by DCGAN to expand the training set, the F1-score of the trained detection model on the same test set is improved from 0.769 to 0.877 after using image fusion to expand the training set, 12.3% increase. The F1-score of the trained detection model on the same test set is improved from 0.769 to 0.877 after using image fusion to expand the training set, 12.3% increase.