Research On Improvement Of Object Detection Model Based On Deep Learning And Its Application In Power Grid

With the development of deep learning,deep convolutional neural network has greatly improved the researches on all of computer vision tasks.As a basic task of computer vision,object detection can predict the classification and localization of object.Since deep learning was introduced in object detection task,many representative researches has been published,which greatly improve the performance of object detection.However,the exist models of object detection and their application precision in actual scene can’t meet the application standard,and further research work is needed.Based on the transmission overhaul application scene,this paper studies the one stage object detection model,and optimizing algorithm and improving performance based on application scene characteristics,to meet the efficiency and high-precision requirements of application.Through delving into the network structure of object detection model,we carry out the optimizations of multi scale object detection,model tasks and feature network to improve object detection performance.Secondly,these improved object detection models are used to detect bird threat intelligently and automatically in power grid overhaul scene.These research works are introduced detailly as follows:(1)An improved multi-scale object detection model based on input is proposed.Bird detection in transmission line scene is a typical multi-scale object detection task.To improve detection performance of small object,an improved SSD-MSN(Multi Scale Object Detection Network Based on SSD)model is proposed based on SSD(Single Shot MultiBox Detector)model.SSD-MSN firstly uses the smaller slices of initial image as model input to enrich the input information of small objects to improve their detection performance.Secondly,in order to reduce the computation of NMS(Non Maximum Supression),SSD-MSN adds an extra area proposal network for the original image,it is used to distinguish the slice contains objects.And the predictions of the slice containing objects and initial image would be integrated as the input of NMS,which can greatly improve NMS efficiency without decreasing detection performance.(2)An improved multi-scale object detection model based on task is proposed.RetinaNet-Conf model is proposed to solve the tasks misalignment problem in RetinaNet model.The main reason of tasks misalignment is that classification performance is inferior to localization performance.An object confidence task is added in RetinaNet,which uses the IoU of prediction box and ground truth box as prediction target,while it shares head network with classification task.The accurate localization task is used to improve classification task to reduce the gap of them,which can solve the misalignment problem of classification and localization tasks.In inference,the product of object confidence and classification score is used to guide the selection process of predicted boxes,so as to balance them and improve detection performance.(3)An improved multi-scale object detection model based on feature network is proposed.In order to raise training efficiency and rich feature,the backbone network of detection model generally transfers from image recognition model pretrained on ImageNet.However,classification task focuses on salient region features of object,while localization task mainly focuses on edge features of object.Therefore,the pre trained backbone network is suitable for classification,but not for localization.To solve this problem,a DSA(Decouple Self-Attention)module based on self-Attention mechanism is proposed.It is located after pretrained backbone network and uses two branches with self-Attention to extract features for different tasks.DSA module and pretrained backbone network can used together,and DSA module based on attention mechanism can extract features containing spatial context information for different tasks,also the experiment results show that DSA can improve detection performance.(4)The applications of improved object detection models in transmission line overhaul scene.Transmission line will be influenced by droppings,nests and stays of birds,so it is necessary to detect and drive away the birds in time.Based on this,an intelligent and all-weather bird detection and prevention scheme based on sound and image is proposed.This paper mainly introduce the bird detection method based on image,including dataset building and the applications of improved detection models.As bird object is small object in transmission line scene image,so this is a small object detection task,and the improved models based on input achieve better detection performances.In order to save calculation,SSD-MSN is selected for bird detection task.And SSD-MSN can gain 83.67%AP,and the precision of bird recognition is 93.2%calculated on detection results,which meets the practical requirements.