Research On Object Detection Of Few-shot In Underground Tunnel Cable Environment

The internal environment of the tunnel is complex,the ventilation is not smooth,dark and humid,which threatens the safety of inspectors.But in order to ensure the power safety and eliminate the hidden dangers,tunnel inspection is essential,so inspection robot has become an important alternative means of tunnel inspection.Fault detection based on machine vision is one of the core technologies of inspection robot.At present,the mainstream method is based on deep learning algorithm.However,the image samples provided by power tunnel are few and difficult to collect.The samples available for deep learning training are seriously insufficient,which limits the research and development of visual algorithm.In this paper,we analyzed the collected cable tunnel video,Aiming at the problem of too few samples of cable equipment in the tunnel environment,the network of target tracking and the basic method of learning with less samples are referred.Based on the Mask R-CNN network and the Siamese network,the method of target detection and segmentation with few samples is studied.Through the network,the pair of image feature information is extracted at the same time,and the similarity is compared.The feature map is sent to Mask R-CNN for subsequent steps to get the detection and segmentation results.Different from the conventional target detection,this paper is equivalent to training a target searcher,that is,given a reference image,to find the target object.In this paper,the method is applied to the detection of tunnel cable.When the target object is not used as the training sample,the target detection and segmentation of one shot is realized by inputting the reference sample.Finally,in the case of limited sample size,a comparative test based on model tuning and information modeling is designed,which further proves that the target detection and segmentation method in this paper has a good detection ability for the equipment under the same sample size.