Study On Insulator Identification And State Detection Method Based On Aerial Image

Electrical insulators as an indispensable equipment of power network, and have the vital significance to maintain its safe operation. However, due to surrounding environment of the power lines is often quite poor, insulator non-destructive aging is serious; however, due to the large number of insulators, wide distribution, so the insulator state detection based on aerial image has important practical significance. In this paper, based on the characteristics of insulators and image processing technology, deep studied of the regional division of aerial insulator image, feature description and target extraction method, and optimized the method of artificial intelligent target detection. The main work is as follows:Different degree evaluation method based on membership function was proposed, and it provided a reference for quantitative measure of the diversity of things. The in-depth analysis of differences in the nature of things, the membership function of fuzzy mathematics is introduced to quantitatively measure the difference, finding the relationship between the degree and the nature of the membership function, and through the analysis of examples, showing that the measure of the difference can more accurately measure differences.There is no actual physical meaning for the looking key point, scale invariant feature transform has high computational complexity, the method based on the difference of the scale invariant feature transform was proposed. According to the description of the difference matrix, the difference values of the feature points were extracted, obtaining the difference matrix. Experimental analysis and results show that the algorithm can effectively match the image, under identical repeatability conditions, effectively reducing the matching error and time complexity.The image region partition method and image feature vector describing method based on the difference value were proposed, the image region partition method used maximum variance between clusters method, it made gray values translate into difference values, according to the difference value had an effective image segmentation; According to the area of image segmentation, the feature points had feature vector description, single regional characteristics description and multi-regional joint characteristics description are proposed respectively, and proved its invariant features.According to the deep restricted Boltzmann machine model of the deep belief network, from the perspective of maximizing the information entropy derivation, level number determine criteria and weight distribution were obtained, according to this criterion and feature, and it can effectively reduce the complexity of the algorithm, and reduce the training time. Aiming at the disadvantage of slow convergence of deep belief networks, the method that using stochastic characteristics of each state and image linear superposition processing was presented, the convergence of deep neural network was sped up, so as to improve the learning efficiency, experimental verification that the time complexity was significantly reduced, and the recognition effect was better.