Research On Few-shot Machine Learning Method In Power Inspection

The power transmission line inspection that is based on UAV onboard line inspection equipment has a number of advantages,including small investments,low costs,automation,intelligence,safety and high efficiency.However,classes of power inspection are numerous and complicated.It's difficult to collect enough training samples for each type.In addition,manual annotation for training samples is too costly.Fortunately,the few-shot classification algorithm is very suitable for this case,in which there are numerous basic classes of training samples but there is no annotation for new classes.However,previous studies seldom focused on this in the field of power inspection,which is a specific business scenario.Therefore,it's necessary to further study the few-shot classification algorithm,make clear its principles and specify its data basis,in order to apply it to the intelligent power inspection system.Based on the neural network models of classical classification,in this paper,the author employed two methods for improvement: the addition of an attention mechanism-based few-shot classification weight generator and change of the classifier of convolution neural network model to the function of the cosine similarity between the feature vector and class vector of the test samples.The methods reserved such an advantage of other classification methods of small samples as better reduction of intra-class variance while achieved better adaptive capability to training sets of higher domain difference.In conclusion,in this paper,the author first summarized the progress and achievements of small-sample learning methods in the past few years,explained the characteristics of and main problems addressed by different algorithms and compared several typical smallsample learning methods,including the matching network,the prototypical network,the relationship network and the model-Agnostic meta-Learning method.Then,the author designed a consistent experimental program to reduce the difference in performance resulted from the complex structure of the few-shot algorithms and the diversified training details,thereby developing a quantitative analysis of the actual performance of each algorithm.After that,the author discussed the training objectives of few-shot learning methods,proved that a deeper backbone network could significantly narrow the gap in performance between different methods and identified through the experiment that infra-class variance was one of the main factors contributing to the difference in performance between small-sample learning methods.Finally,the experiment results also demonstrated that,as the domain difference increased,reinforcement of the adaptive capability to a few new-class instances could better enhance the accuracy of the algorithm compared with reducing infra-class variance.The results also served as a reference to the preparation of data sets in the field of power inspection.