Research On Lightweight Object Detection Algorithm Based On Deep Learning

Object detection is an important task in the field of computer vision,which refers to the technology of detecting and extracting specified targets from images or videos.This technology is widely used in autonomous driving,big data for transportation,intelligent security and industrial testing,etc.The related research of object detection technology based on deep learning has achieved certain results.However,there are still some problems and challenges,one of which is the challenge of network lightweight.The object detection networks based on deep learning often takes accuracy as the priority principle in design,pursing high-precision detection of multiple types of objects.As a result,such a network is bulky,which is difficult to deploy on edge computing platforms.Besides,some small models have the problem of low recognition accuracy.In the application scenarios of most object detection models such as face recognition and text detection,there are fewer target categories to be detected,and there is a high demand for real-time computing performance.Therefore,while balancing the network accuracy,compressing the network volume and designing a universal lightweight object detection network is of great significance to the application of deep-learning-based object detection technology.Aiming at the above problems,the main research contents of this thesis include:1.Aiming at the problem that the current object detection networks have large amount of parameters and large model volume,which can not be deployed on the edge computing platforms,the YOLOv5-lw network,which is lightweight improved based on YOLOv5,is proposed.The network uses a new backbone,which is composed of Ghost Bottleneck,to extract features of images.Besides,the network is improved by using lightweight technologies such as asymmetric convolution and depth separable convolution.The experimental results show that the parameters and model volume of YOLOv5-lw are significantly reduced,and the detection speed on PASCAL VOC data set is significantly improved compared with the traditional object detection network,with little loss of accuracy.2.Aiming at the problem of accuracy loss caused by the lightweight technologies,YOLOv5-lwa network,which is improved based on YOLOv5-lw by introducing attention mechanism and optimized feature fusion module,is proposed.A channel based attention module is added to the backbone of the YOLOv5-lw network,and the concatenation module is replaced by FFM(Feature Fusion Module)to improve the learning and feature fusion ability of the network.The experimental results on PASCAL VOC data set show that the YOLOv5-lwa network maintains a small volume and improves the accuracy compared with YOLOv5-lw network.3.After proposing the lightweight object detection network YOLOv5-lwa,the network is further applied in the field of intelligent driving.Aiming at the problem that the accuracy of the current Chinese guide sign text detection algorithm is not high and cannot perform detecting in real time,a lightweight Chinese guide sign text detection and recognition algorithm,which is improved based on YOLOv5-lwa network,is proposed.The algorithm uses YOLOv5-lwa network to detect the text areas,and uses a new text split algorithm based on projection histogram to split the text areas into single characters.Finally the lightweight classification network is used to classify and recognize the characters.The comparative experimental results on the TS-Detect guide sign data set show that the algorithm can recognize the text information in guide signs in real time under complex shooting conditions and has high accuracy.