Research And Implementation Of Network Device Backplane Detection Method Based On Deep Learning

Network device backplane detection refers to the use of object detection technology to detect the backplane of network devices such as switches and routers.It can identify the main body and main components of the backplane in the image,so as to automatically obtain the key information such as the composition structure,bandwidth capacity and operation status of the backplane.Hence,backplane detection can significantly improve the efficiency of the operation and maintenance management of the network devices.The current two-stage object detection model based on deep learning is mainly Faster R-CNN(region based convolutional neural networks).It uses RPN(Region Proposal Networks)to predict anchor's binary classification scores and bounds to generate region proposals and then extracts features from region proposals to predict classification scores and bounds to generate detection boxes.There are two main problems of Faster R-CNN in the backplane detection scenario:there are large differences in object scales on the backplane,and the anchor scales generated by the existing anchor generation mechanism can not guarantee a high consistency with all object scales;the netports on the backplane are densely packed,and the existing NMS(Non-maximum Suppression)algorithm can not effectively remove the duplicated detection boxes of the netports.In this thesis,based on improved Faster R-CNN,the backplane detection method considering the characteristics of the detection scenario is proposed.This method proposes an anchor generation mechanism to take effect in RPN and proposes the deduplication mechanism of dense netports to take effect in the post-processing stage,so as to increase the accuracy of improved Faster R-CNN.The model could realize the precise detection of the backplane board,network cable port,optical fiber port,USB port,status indicator,and manufacturer.The work of this thesis consists of two parts.Firstly,the backplane detection method based on improved Faster R-CNN is proposed.On one hand,This method proposes the class-specific k-means anchor generation mechanism to improve the consistency between anchor scales and object scales;On the other hand,due to the limitation of the NMS algorithm in the dense netport scenes,the deduplication mechanism of the dense netports is proposed to achieve the effective deduplication of the detection boxes of netports.The experimental results show that the mAP(mean average precision)of the improved detection model reaches 90.3%,which is 6.7%higher than that of basic Faster R-CNN.Secondly,based on the proposed detection method,this paper designed and implemented the backplane detection and vectorgraph conversion system,which can build the backplane dataset for training and test,as well as the construction and update of detection models,and provide the webservice of backplane detection and vectorgraph conversion.