Research On Object Detection Technology Based On Lightweight Model Compression

In recent years,with the rapid success of convolutional neural networks(CNNs)in fields such as image processing,autonomous driving,and semantic segmentation,the pursuit of model accuracy has led to an exponential growth in the number of parameters and computations required for these models.As a result,the deployment of large deep learning models on low-powered devices such as mobile and embedded devices has become increasingly challenging,as lightweight models tend to run slower.Therefore,it has become crucial to use model compression algorithms to reduce parameter redundancy and lower storage costs.In response to existing problems with excessive model parameters and slow operation speed of lightweight models,this paper proposes a model compression algorithm that combines pruning and knowledge distillation.This algorithm reduces the number of model parameters,improves inference speed,and achieves deployment on embedded devices such as the Jetson Nano to detect vehicles.The research work of this paper is summarized as follows:(1)Introducing channel pruning algorithm based on YOLOv4-tiny model.The YOLOv4-tiny model has a large number of parameters,leading to slow model operation.To address this issue,we introduce channel pruning to calculate the rank of each feature map matrix and determine the importance of each channel to the network model.By pruning the YOLOv4-tiny model,we can reduce the number of parameters and eliminate redundancy.(2)Introducing knowledge distillation algorithm based on YOLOv4-tiny model.Although pruning improves detection speed,it may result in reduced detection accuracy.To address this problem,we sample local data points near the target object in an image,and use knowledge distillation to extract sample information from a high-accuracy teacher model and enable a student model to learn from it.This helps to continuously improve the detection accuracy of the pruned network model.(3)Implementing the deployment of the YOLOv4-tiny network model on the Jetson Nano development board for vehicle detection.To verify the feasibility of the compressed network model,we deployed the original model and the compressed model on the Jetson Nano development board and compared the frame rate of the two models.Finally,we achieved smooth operation of the compressed model on the Jetson Nano development board to detect vehicles.Through experimental design and comparison with major model pruning algorithms in recent years,our results show that our method can reduce the number of model parameters to only 2.87 M while improving FPS by 93.7%,with a detection accuracy loss of only 2.9%.