Design And Implementation Of Lightweight Object Detection Model

By the support of large-scale public datasets and increasing hardware computing power,deep learning has achieved great success in many fields of computer vision.Models based on convolutional neural networks have become the mainstream solutions in object detection.Researches on object detection can be divided into two categories.One is high-precision model,which aims to refresh the precision records of public datasets using equipment with high computing capacity,the other is lightweight model,which focuses on lightweight and real-time performance with trade-offs between accurate and lightweight,which makes it more suitable for practical applications.Most of current lightweight techniques only focus on certain steps,and lack the guiding ideology and a unified framework throughout the entire implementation process.This thesis studies the generic lightweight object detector implementation framework.In order to balance accuracy and weight of the model,this thesis studies on construction of a generic lightweight network and enhancement of specificity of the network.First,this thesis combeds the constraint conditions of detection tasks and complexity indicators of the network structure.Three structural factors,depth,width,resolution of the network can directly affect performance.These conditions and indicators are considered synthetically using roofline model to propose disciplines to build a lightweight detection network.Next,this thesis introduces one-shot aggregation(OSA)connection and cross-stage partial network(CSP)to improve backbone of mainstream detector,and uses CSPOSA modules to construct a generic network.Since there is a certain redundancy for a specific task in generic networks,this thesis optimizes the model based on the network structure and the whole training process.The width and depth of the network are adjusted to adapt the specific task and the weight of the model is furtherly compressed.The training process is divided into three stages to optimize respectively.This thesis selects two actual applications,which are security helmet detection deployed on embedded systems,and pedestrian/ vehicle detection deployed on edge AI modules,as experiments to verify the effectivity of the framework.Weights of the models implemented using the framework is 1/10 ? 1/100 of mainstream models.Models implemented by the framework proposed in this thesis are more friendly to deploy on devices with limited computing power.