Research On Lightweight Object Detection Algorithm Based On Deep Learning

The thesis aims to tackle the issue of high computational complexity in mainstream object detection algorithms by conducting research on lightweight object detection algorithms based on deep learning,with a focus on the YOLOv7 one-stage object detection framework.As the parameter size of deep learning-based object detection algorithms is showing exponential growth,which limits their deployment on resourceconstrained edge devices,the thesis explores strategies such as compact network design,network pruning,and knowledge distillation to develop more lightweight object detection models.The main research contributions of the thesis are as follows:(1)A lightweight object detection network model based on YOLOv7 and Mobile Net V3 is proposed to address the problem of high complexity in the YOLOv7 algorithm model.A newly designed Lightweight Mixed Attention Module(LMAM)is used to replace the attention mechanism module in the Mobile Net V3 model and the core module of the Mobile Net V3 model is restructured.Finally,the improved lightweight model is used to replace the backbone network of YOLOv7,resulting in a more lightweight and feature-rich object detection model,YOLOv7-M.On the VOC public dataset,the computational complexity of this lightweight model is reduced by 7.24 GFLOPs compared to YOLOv5-tiny and 3.94 GFLOPs compared to YOLOv7-tiny,and the model accuracy is increased by 2.35% compared to YOLOv5-tiny and 1.2% compared to YOLOv7-tiny.(2)An automatic network pruning algorithm based on attention mechanism is proposed to address the problem of parameter redundancy in the object detection model.Using attention modules to quantitatively analyze the importance of each convolution channel within a convolutional layer,combined with an improved particle swarm algorithm,the original model is progressively channel-pruned from a global perspective to achieve sparse optimization of the object detection model while maximizing the model accuracy.Finally,the effectiveness of the automatic pruning algorithm is verified on multiple public datasets.(3)A knowledge distillation algorithm suitable for object detection scenarios is proposed to address the problem of a sharp reduction in model accuracy at a high pruning rate.By fusing the feature maps of the teacher and student models in the backbone network and decoupling foreground and background regions of the image in the head network,the student model can better learn the hidden feature information in the teacher model.The effectiveness of the knowledge distillation algorithm is verified on multiple public datasets.In summary,based on methods such as compact network design,network pruning,and knowledge distillation,the thesis proposes a solution for lightweight object detection algorithms based on deep learning,and uses a one-stage object detection algorithm as an example to ultimately obtain a lightweight object detection model with a parameter size of 3.28 M,a model accuracy of 74.58%,and a processing frame rate of 205 FPS.The innovative methods proposed in the thesis provide insights for developing lightweight object detection algorithms.