Object Detection Based On Lightweight Convolutional Neural Network

In recent years,with the rapid development of computer vision,Convolutional Neural Network(CNN)has been widely used in fields,such as natural language processing and medical image understanding.As one of the basic tasks of computer vision,object detection has important research significance in the fields of autopilot,face recognition,and scene understanding.Lightweight Convolutional Neural Network is an application-oriented CNN network architecture design,and is the research basis for various computer vision tasks applied to low-power devices.In this paper,based on the efficient feature extraction capability of the densely connected network and the lightweight structure of the depth-wise separable convolution,the densely connected lightweight feature extraction network CED-Net(Channel Enhanced Dense Connection Network)is first proposed.Then,this paper proposes a dense feature pyramid structure that combines shallow and deep feature maps.Finally,combining the advantages of two networks,this paper designs a lightweight object detection model named CED-De.The main work of this paper is summarized as follows:First,this paper propose a lightweight convolutional neural network based on dense connections.By analyzing the principle of the convolutional layer and lightweight architecture,the bottleneck layer convolution module and the downsampling module based on the depth-wise separable convolution kernel are designed.In order to improve the problem of loss funcation of feature information caused by depth-wise separable convolution.By cite the channel enhancement module composed of squeeze-and-excitation blocks to enhance the network's intermediate feature map channels and improve the ability to express feature information.On the basis of this work,a lightweight convolutional neural network architecture CED-Net is designed.The CED-Net performed better on CIFAR and ImageNet datasets.Compared with other convolutional neural networks,it performed better in inference speed.Then,a dense feature pyramid network and a lightweight object detection model CED-Det are proposed.The dense feature pyramid network can increase the representation information of the network output feature map by fusing shallow and deep features.Based on SSD object detection network,combined with CED-Net and dense feature pyramid network,designed a lightweight object detection model—CEDDet.This model first extracts features from CED-Net,and then performs feature fusion by stacking two layers of dense pyramid networks.Finally,the feature maps after fusion are predicted by two 3 × 3 convolutions.The experimental results of CED-Det on the VOC and COCO datasets show that compared with other objectdetection models,CEDDet is more suitable for embedded platforms in terms of accuracy,inference speed,and total parameters.Finally,a stop line tracking system based on a lightweight object detection model is designed.The object detection task is applied in autopilot,using the image data collected by actual vehicles on urban roads as the data set,and the CED-De is trained to detect the stop line,and the continuous tracking is achieved by the SORT object tracking algorithm.Stop line tracking in video frames.This paper has 33 figures,8 tables,and 106 references.