Research On Application Of Object Detection Based On Deep Convolutional Neural Network

Object detection shows substantial performance boost for efficient CNNs that the incapability to meet all scenarios.For example,moving vehicles and pedestrians,as the part of the traffic,easily break the traffic order.Detection of vehicles and pedestrians is conducive to a smart city.For example,insulators are parts of overhead transmission lines,which bear extremely high mechanical strength.Self-explosive insulators damage the normal operation of transmission lines.As a consequence,Self-explosive insulators must be replaced.Artificial intelligence empowers industry development.As its basic subject,object detection faces challenges in the application of complex scenes.In the context of deep learning see the world,we improved the object detection algorithms and applied them to the detection of urban traffic participants,the detection of multiple license plates under complex backgrounds,and the segmentation and detection of self-explosive insulators in China Southern Power Grid.1)In autonomous driving scenarios,wrong object detection leads to serious accidents.We proposed an object detection embedded with attention and feature interleaving modules.This method mainly improves the channel of the feature map: Firstly,the attention module learns the weight of each channel,enhances key features and suppresses redundant features,thereby strengthening the network’s separation of objects and background;at the same time,the channels are grouped and interleaved to obtain more representative features;finally,the features obtained by the two modules are merged into the next layer.2)In order to detect multiple license plates with different colors and sizes in a complex background,we propose an improved YOLOv3,which takes into account accuracy,network complexity and detection time.The main improvements include,firstly,integrating the Inception-Res Net module into the backbone network;secondly,adding the SPP module to the detection head;then,clustering the license plate size and designing anchors;finally,cutting the repeated modules to reduce the training parameters and detection time.3)For insulator string segmentation and self-explosive insulator location,we propose a segmentation model of insulator string and a detection model of self-explosive insulator.Through intimacy modeling and contour modeling to learn the appearance commonality and contour commonness of insulators,the network can perceive the integrity of the foreground instance and the separability from the background.Then the self-explosive insulator position is detected from the mask map of the insulator string.