Research On Image Semantic Segmentation Technology Of Substation Road Scene Based On Convolution Neural Network

With the rapid development of power system and artificial intelligence,substation inspection robots are gradually replacing manual inspections.During the inspection process,if the visual information can be used to assist the navigation system of the inspection robot to understand and analyze the environment in which the robot is located,the reliability of the navigation system of the inspection robot can be further improved.Therefore,from the perspective of two dimensional image scene understanding,this thesis focuses on the semantic segmentation technology of substation road scene image based on convolutional neural network,and assists navigation system of inspection robot by using visual information appeared in the field of robot's view.The main contributions are as follows:The adaptability of various convolutional neural network structures to substation road scenes is explored.In the process,a semantic segmentation data set of the substation road scene image is generated,which includes six types of objects: road,stone,grass,fence,pit and background.From the practical point of view,a number of evaluation indicators are determined for evaluating the performance of the relevant semantic segmentation network.Based on the classification network,the full convolutional neural network based on VGG,Res Net and Squeeze Net are designed to realize the semantic segmentation task of substation road scene image.After the evaluation,a basic semantic segmentation network called Squeeze Net-8S is selected.According to the selected basic semantic segmentation network Squeeze Net-8S,its structure is improved from its branch convolution layer in infrastructure and the fusion of multiple layer's output feature maps.In terms of convolution,combined with the ideas of Dilation and Xception structure,two improved structure called D-fire-module and XD-fire-module are proposed.Then,a D-Squeeze Net-8S network and a XD-Squeeze Net-8S network are built according to the improved structure.In the test,it is found that for the substation road scene,the D-Squeeze Net-8S network has better semantic segmentation accuracy than Squeeze Net-8S while ensuring the same quantity of network parameter,and the semantic segmentation accuracy of XD-Squeeze Net-8S network is slightly improved on the basis of the D-Squeeze Net-8S network,but its parameter of network is smaller than D-Squeeze Net-8S network.In the aspect of multiple feature map fusion,an ASPP structure is introduced in the skip layer to improve the performance of network.An AXD-Squeeze Net-8S network is designed for the improved structure.The test result shows that the semantic segmentation accuracy of the improved network is better than XD-Squeeze Net-8S,but the introduction of ASPPstructure also increases the number of parameters of the whole network.In summary,this thesis explored the adaptability of multiple convolutional neural network structures for substation road scenarios.Based on this research,a full convolutional neural network called Squeeze Net-8S is selected for this sense.The scene image is semantically segmented,and the basic structure of Squeeze Net-8S is further improved from its branch convolution layer of infrastructure and multiple feature map fusion in skip layer.D-Squeeze Net-8S,XD-Squeeze Net-8S and AXD-Squeeze Net-8S networks are built according to the improved related structure.Then,these improved semantic segmentation networks are used to realize the semantic segmentation task ofsubstation road scene images,combined with the semantic segmentation result of network's output,the corresponding post-processing work is carried out by digital image processing method,and finally this thesis realized the task of providing visually information to assist the decision of robot's navigation system.