Development Of Lightweight Object Size Adaptive Detection Algorithm

Object detection is a fundamental technology in computer vision.Recently,object detection in deep learning has a rapid development because of the great success achieved in image classification.But it's difficult to have good effect on dense object detection because of small objects and occlusion.Also,the corelated algorithms aim at upgrading precision on benchmark,so they run slowly and have massive parameters and FLOPs,As a conclusion,they don't have good engineering application value.So we develop a lightweight network for detection.And we develop object size adaptive algorithms for dense scence.Our work has good engineering application value.Firstly,we develop the lightweight detection network for object detection.Through the development of architecture,we get the network with lower parameters,lower FLOPs,faster inference speed.And then,we develop the pruning algorithms to get the sparse lightweight network which could get lower calculate consumption with lower loss of mAP.Knowledge distillation algorithms are developed to get higher mAP while maintain the same architecture.At last,the algorithms for dense object detection are proposed.We firstly collect the dense car dataset.Afterwards,we propose the brand new pre-and post-processing methods.And we propose a generic loss function to solve regression problem.Results show that the compressed sparse light weight network could get 96%lower parameters,98%lower FLOPs compared to base network.And could run at 76fps on Telsa P4.Our algorithms could get 86.1%mAP and 86.6%Recall on dense car dataset.When compared to our base algorithm,we could get 6.6%higher mAP and 4.5%higher Recall.