Research On Semantic Segmentation Method Of Key Components Of Aerial Transmission Line Based On FCN

The operation status of many power components in the transmission line directly affects the transmission safety.Semantic segmentation of the aerial image can provide a technical basis for other advanced visual tasks in the later stage.This thesis takes aerial image of the transmission line as the research object.Aiming at the current research status and existing problems of the key component segmentation of aerial image of transmission line,the thesis applies the deep learning theory and completes the following work:In the process of semantic segmentation of transmission line aerial image by FCN algorithm,it is necessary to construct a transmission line aerial image semantic segmentation dataset to realize the training of deep model.In order to solve the problem that it can't train the deep model without dataset,the thesis constructed a semantic segmentation dataset of transmission line aerial image,studied the annotation format and methods of related semantic segmentation datasets such as PASCAL VOC,MS COCO and Cityscapes,and completed the annotation of the dataset with the annotation method of the semi-interactive GrubCut algorithm.The key target material in the aerial image of the transmission line,the change of illumination and other factors will cause deviation when calculating the feature,affecting the accuracy of the traditional segmentation method.In order to solve the complexity of the traditional segmentation algorithm and the artificial selection feature has low universal applicability to different types of targets,a method based on FCN for semantic segmentation of aerial image of transmission line is proposed.The migration learning method is used to train the deep network.Experiments show that FCN can realize end-to-end semantic segmentation of key targets of transmission lines.Aiming at the problems of poor segmentation accuracy and missing edges of key targets in aerial image of transmission line based on FCN model,multi-scale expansion convolution is adopted in the up-sampling stage of extracting features from deep network and probability graph model is connected to optimize the segmentation performance.Experiments show that the improved FCN improves the average pixel classification accuracy by 2.43% and the average crossover ratio by 3.58%.