| Object detection is a fundamental problem in the field of computer vision,which involves detecting relevant objects in images.It has significant practical value and theoretical significance in fields such as autonomous driving,surveillance,robotics,and medical imaging.However,since object detection models are typically deployed on low-power edge device platforms,many existing high-accuracy large models are not suitable.As a result,designing an appropriate object detection network architecture is a heavily experience-dependent task.Therefore,it is very necessary to study how to automate the design of high-performance and low-power object detection network architectures.Neural Architecture Search(NAS)utilizes intelligent algorithms to automatically search for network architectures to eliminate the dependence on expert experience.It is an effective network optimization and pruning technology.However,most NAS algorithms are mainly focused on image classification tasks,and NAS for object detection is still in the initial research stage.(1)In order to adapt to the fine-grained requirements of object detection models in different deployment scenarios,this paper proposes a neural architecture search algorithm for object detection models based on the Big NAS algorithm for classification tasks.Firstly,a searchable reparameterized multi-branch unit is proposed,and an intergral search space(supernet)for object detection is constructed,including the backbone network,feature fusion network,and upward network.The search space includes the size of convolutional kernels,the width and depth of the model,and the connection method between the fusion network and the upward network.Then,using the sandwich rule and local self-distillation strategy,the weights of the supernet are fully trained to evaluate the accuracy of the subnetwork later.Finally,a evolution search strategy based on group sampling based on model computational complexity is proposed,which can explore the subnetwork models more uniformly and efficiently in the search space.Compared with classical object detection networks,the excellent subnetworks searched with lower computational costs maintain similar accuracy.(2)In response to the need for detecting fire risks in construction sites,this paper designs and implements a site fire hazard detection system for construction sites.Firstly,by collecting real cutting and welding operation scene image data and labeling it,a dataset necessary for the detection system can be obtained.Then,based on transfer learning technology,the object detection supernet model is fine-tuned using the collected dataset to obtain the optimal subnetwork model that can be deployed on edge devices.Finally,a web-based software interaction system is constructed,and highperformance searched subnetwork models are accelerated and deployed on the NVIDIA edge computing platform Jetson AGX Xavier.In summary,this study aims to automate the design of high-performance and lowpower target detection network architectures using neural architecture search(NAS)technology.In addition,based on the practical application needs,this paper designs and implements a fire hazard detection system for construction sites,which demonstrates the effectiveness of the proposed algorithm. |
