Infrared Image Combination Expansion Method Based on Supervised Single Sample Power Equipment

Due to the particularity and confidentiality of the power industry, the infrared images of power equipment do not have a good public data set to use, and the collected sample data information is extremely unbalanced, and too little data can easily lead to overfitting of the network. In order to solve this problem, this paper finds a most suitable augmentation method for infrared images of power equipment based on the supervised single-sample image augmentation method. Firstly, the characteristics of common augmentation methods are analyzed, and then the five methods of noise, flipping, random cropping, scaling and contrast in the single-sample augmentation method are selected for detailed analysis according to the unique characteristics of infrared images of power equipment, and finally these five methods are verified based on VGG16 and ResNet50 classification models. Case analysis shows that the contrast, flipping and noise expansion methods have better expansion effects, and the effect of the double expansion combination is better than any single expansion method under the same amplification increment, among which contrast + flip and contrast + noise are the best combinations of the double combination methods, and the detection accuracy is improved by 6.17%, 5.90%, 0.98% and 0.69%, respectively.