Research And Implementation Of Lightweight Object Detection Algorithm Based On Deep Learning

Object detection is an important research direction in the field of computer vision,which is widely used in many fields such as transportation and medical treatment.At present,the object detection algorithm based on deep learning has made a breakthrough in accuracy,but it still cannot be widely used in embedded devices with limited computing and storage capability due to its large number of parameters and calculations.This thesis proposes a lightweight object detection algorithm based on SSD,which is obtained by four parts:network optimization,channel pruning,knowledge distillation and storage optimization.The details are as follows:(1)In the network optimization part,in order to solve the problem of insufficient feature extraction ability and large parameter amount of standard SSD network,this thesis designs an optimized SSD network based on DRN22.(2)In the channel pruning part,aiming at the difficulty that the existing methods are difficult to maintain high accuracy when the pruning proportion is large,this thesis proposes a mixed pruning method which combines the automatic determination method and manual definition method.Meanwhile,in view of the automatic determination method,a global pruning standard/~2 based on theandparameters of BN(Batch Normalization)layer is proposed.In view of the manual definition method,a local pruning standard based on the intersection of the L2 norm and the geometric median is proposed.(3)In the knowledge distillation part,this thesis designs to transfer knowledge by constructing attention map loss,weighted KL divergence loss and teacher bounded regression loss,which alleviates the accuracy reduction problem caused by channel pruning.(4)In the storage optimization part,the storage volume of the network is further compressed by merging BN layer,INT8 quantization and Huffman coding.In addition,based on the network,this thesis completes the design and implementation of the embedded system on the Jetson TX2 platform.The experimental results show that:The lightweight object detection network designed in this thesis has a average precision of 75.1%on the PASCAL VOC dataset,whose parameter amount is only 4.4M and inference speed is 106FPS.It has a good balance between accuracy and speed.The inference speed of the embedded system based on the lightweight object detection network is 25.3FPS on the Jetson TX2 platform,which has a good landing value.Overall,the lightweight object detection network in this thesis has reached the expected targets.