Object Detection Based On Lightweight SSD

Object detection based on deep learning is one of the most crucial techniques in the field of computer vision,and has been widely applied to transportation,medical treatment,military and other areas.In recent years,scholars have proposed many Convolutional Neural Network(hereafter as CNN)structures successively with superior performance,which also improved the performances of object detection models.However,this improvement always comes with a multiplicative increase of the network scale.Especially in the tasks requiring a high level of real-time performances,the network scale is greatly limited by the storage space and computing power of the device.To solve the problems above,this dissertation proposes a design of lightweight CNN structure and a lightweight object detection model based on SSD network.The study results are as follows:(1)This dissertation propose a lightweight Convolutional Neural Network(Multi-size Feature Map Augmentation Network,MFA-Net)based on feature map augmentation.Firstly,in order to combine the advantages of large size convolution kernel without introducing too many parameters,different size convolution kernels are introduced into the same depthwise convolution and Multi-size depthwise convolution is proposed.Secondly,to study the redundancy of feature map,the author replaces the traditional convolution that causes redundancy with multi-scale deep convolution,in order to decrease the parameter quantities during the generation of multi-scale feature map.Under the guidance of lightweight principle,a feature map augmentation module is constructed.Finally,a feature augmentation convolutional block is constructed based on the features of residual block and the advantages of the feature map augmentation module.On that basis,the author put forward with lightweight CNN,which is MFA-Net.The network displays fine classification performances in image classification datasets CIFAR-10 and Image Net-1K,and also has an edge over other CNN in terms of network scale and computing speed.(2)A lightweight object detection model based on SSD,Light-SSD is proposed.Firstly,the author replaces the VGG-16 network in the original SSD with the lightweight CNN mentioned above,MFA-Net.Considering the network scale,the author introduces a lightweight attention module to make up for the performance loss of lightweight network.In addition,dilated convolution is also introduced to optimize the feature map augmentation module and further broaden its receptive field.Therefore,a lightweight target detection model Light-SSD is designed.Finally,the author adopts augmentation strategies like Cutmix,Batch Group Normalization,multi-scale training,and generalized focal loss in model training,to further improve the model performance.In the VOC and COCO object detection tasks,the Light-SSD model is more suitable for embedded platform in terms of parameters and inference speed,and still maintains a high accuracy.(3)A design of a real-time mask detection system based on Light-SSD model is proposed.Firstly,to complete the real-time mask detection tasks,the author collects and constructs Masked Face Extended Dataset.Next,a real-time mask detection system based on computer vision library Open CV is realized and deployed into the embedded platform Jetson Nano,achieving great detection results.There are 38 figures,7 tables and 81 references in this dissertation.