Lightweight Method Of Object Detection Networks Based On Deep Learning

The massive data generated by the Internet and rapid improvement of computing power greatly promote deep learning,makes object detection a breakthrough.However,Too many layers of deep neural networks lead to a large number of parameters and FLOPS.The application of deep learning depends on high-performance GPU.The High requirements make deep learning difficult to deploy on the low-performance computer.However,due to the cost,many applictions need to do it.To slove these problems,three effective model lightweight methods are proposed,which can reduce the model parameters and FLOPS,speed up the model,and improve the accuray.(1)The Yolov3 network has a lot of parameters and FLOPS,and it runs slow.This paper uses channel pruning and layer pruning to lighten yolov3.Specifically,channel pruning is used to reduce the width of the model,and layer pruning is used to reduce the depth.After using the appropriate purning ratio,a more compact model can be obtained.The trimmed model can resume the accuracy by fine tuning.Experiments are put on the face mask detection task and helmet detection task,then we get two compact models with less parameters and FLOPS,while the detection accuracy is only slightly droped.(2)To improve the precision of the model after channel pruning and layer pruning,knowledge distillation is proposed to guide the fine-tuning process.The model without pruning is taken as the teacher network,and the pruned one is as the student network.During the fine-tuning process,the soft labels output by the teacher are used to provide more information to the student network,which improve the accuracy of the latter.Experiments show the accuracy is improved after pruning.(3)To solve the shortcomings such as a large number of parameters,a large number of FLOPS and poor detection accuracy in the Pelee object detection network,an improved version named GCPelee is proposed based on grouped convolution and feature maps cascade.Firstly,the amount of model parameters and FLOPS is reduced by replacing normal convolution in the detection module with group convolution;secondly,feature maps cascade is applied on the detection module to transmitted the information contained in the feature maps with a large receptive field to the feature maps with a small one,which enlarges the receptive field of the latter.The experiments show that the GCPelee model gets higher detection accuracy with less parameters and less FLOPS.