Research On Target Detection Algorithm Based On Deep Learning

Target detection is a hot research content in the field of photoelectric image processing,and has been extensively studied in the past few decades.The purpose of target detection is to find objects with a specific target class with precise positioning in a given image,and to extract specific information of the target.Since deep learning theory has achieved very good results in image classification and other applications,in recent years,more and more attention has been paid to how to apply deep learning to target detection.However,because deep learning training and reasoning require huge computing power and high hardware cost,how to design a lightweight neural network whose detection effect can be equivalent to a large network model has become a hotspot of current research.At the same time,the need to run more efficient deep neural networks on embedded systems has also promoted research on lightweight neural networks.This article has conducted in-depth research on the design of lightweight network structures.The specific research contents are as follows.This article first introduces the research status and basic principles of target detection algorithms based on deep learning.By analyzing the algorithm complexity of the convolutional neural network,the factors affecting the speed of the convolutional neural network are analyzed from the time complexity,space complexity and memory access cost.Then on the basis of experiments,the research put forward and verified four practical suggestions for efficient and lightweight network design.Equal channel width minimizes memory access cost;Excessive group convolution increases MAC;Network fragmentation reduces degree of parallelism;Element-wise operations are non-negligible.Then for the application requirements of embedded devices with limited computing power,according to the above suggestions,a lightweight convolutional neural network architecture ShuffleNetV2 with high computing efficiency was designed.The network model uses depthwise convolution and random channel shuffle to ensure the detection accuracy while greatly reducing the computational complexity.At the same time,in order to take into account the running speed,the YOLOv3 detection module is used as the detection network backend.While keeping the complexity of the network model unchanged,network distillation is used to further improve the detection performance of the lightweight network model ShuffleNetV2.Use distillation losses to transfer knowledge of a more accurate teacher network to the ShuffleNetV2 lightweight student network.Experimental results show that through feature map non-maximum suppression algorithm and objectness scaled loss strategy,network distillation effectively improves the performance of the lightweight network model target detection algorithm.Finally,train and test the algorithm performance on the self-built train water inlet and ship data set,and the classic target detection data set PASCAL VOC.According to the characteristics of the data set target,the detection module is improved accordingly.The experimental results show that the correctness of the proposed lightweight network structure design suggestions proposed in this paper and the ShuffleNetV2 lightweight target detection algorithm have high detection performance,and the detection speed is greatly improved without a small loss of detection accuracy.At the same time,performing network distillation on the lightweight network model ShuffleNetV2 can effectively improve the detection performance of the lightweight network model and maintain the original detection speed.