Lightweight Research And Application Of YOLOv5 Target Detection Networ

With its powerful learning capability and universality,deep learning has demonstrated outstanding performance in various fields,such as target detection.Although YOLOv5,as a representative model in the field of deep learning target detection,has achieved remarkable improvement in detection accuracy,its complexity is high,making it difficult to be applied to platforms with limited storage and computing resources,such as mobile and embedded devices.To address the above problems,this paper uses YOLOv5 s as the base model for its lightweight research,and the main research contents are as follows:(1)In order to reduce the number of parameters and computation of the model while ensuring the detection accuracy of YOLOv5 s model,this paper proposes a lightweight target detection algorithm GMFF-YOLO based on multi-scale feature fusion.Firstly,this algorithm designs the MFF Bottleneck structure of multi-scale feature fusion and replaces the Bottleneck structure in PANet,so that it can obtain more abundant semantic information and enhance the feature fusion capability of the network.Secondly,a lightweight GSConv structure is used to replace the individual convolutional layers in the YOLOv5 s model,making it more lightweight.Finally,experiments are conducted on multiple datasets to verify the performance of the GMFF-YOLO model.(2)In order to further improve the inference speed of GMFF-YOLO model and make it more lightweight,a model compression algorithm based on hybrid structural channel pruning is proposed in this paper.This algorithm adopts a channel pruning algorithm framework based on BN layer sparsity.Firstly,a sparse factor adjustment strategy is designed during the sparsity training process of the GMFF-YOLO model.Secondly,channel pruning algorithms are designed for the GSConv and MFF Bottleneck hybrid structures of the GMFF-YOLO model,and the GMFF-YOLO model at the end of sparse training is pruned.Thirdly,in response to the problem of decreased accuracy in the GMFF-YOLO pruning model,a knowledge distillation method is used to fine tune its accuracy for training.Finally,the effectiveness of the model compression algorithm is demonstrated experimentally.(3)By deploying the YOLOv5 s,GMFF-YOLO,and GMFF-YOLO pruning models on the Jetson TX2 embedded platform using the Tensor RT deep learning acceleration framework,respectively,and conducting inference experiments using the Tensor RT FP32 model and Tensor RT FP16 model,further verifying the effectiveness and feasibility of the proposed algorithms in(1)and(2)for the lightweight YOLOv5 s model.