Research And Application Of Lightweight Object Detection Based On Deep Learning

In recent years,with the combination of deep learning and the Internet of Things,people have gradually begun to experience the convenience brought by artificial intelligence,and the demand for smart devices has been increasing.Although the application of artificial intelligence technology in the Internet of Things promotes the intelligence of the Internet of Things and makes it more valuable,the artificial intelligence technologies such as deep learning are all too complicated,and the memory and computing resources consumed are very large,difficult to deploy in Io T devices.For this reason,this thesis swill research on the compression and application of the model,aiming to reduce the size and processing time of the model while maintaining the performance of the model.The main content of this thesis are as follows:(1)Based on the analysis of the existing object detection algorithm and network structure,we reconstructed the structure of the YOLOv3-tiny by group convolution and shortcut ideas to reduce model parameters and calculations.Finally,the model size was reduced from 33.26 MB to 3.47 MB,compressed 10 times,the processing time on the embedded platform was accelerated 7 times,and the average accuracy rate was increased by6%.(2)According to the analysis of model parameters,it is found that some of them are inactive or low response state.In this thesis,we proposed a pruning method for redundant or similar channels and a pruning strategy for secondary channel pruning.The model size was further compressed to 1.79 MB,and the average accuracy rate of the model was improved by2% compared to YOLOv3-tiny.In addition,we combined with Neon to accelerate convolution operations,the detection speed reached 0.5 seconds / frame.(3)An edge intelligence system is designed for weak computing servers and embedded devices.The CNN model is divided into two parts,one part runs on the embedded device and the other part runs on the server.This method saves bandwidth while the accuracy is not lossed,it is very useful for edge intelligence system.We take it used in intelligent monitoring system,and the accuracy rate of system is 84%.In summary,for the problem that the deep learning model is too redundant and complicated,we improved network structure and channel pruning method to make model simpler and more effective.And,it achieved fast speed on embedded devices and has certain practical value.