Research On Open World Object Detection Method Based On Robust Optimization

Conventional object detection methods are usually set in the assumption of a closed world,train on specific data sets,learn a fixed number of object categories,and apply them to specific scenarios.With the development of the information age,the speed of product upgrading is accelerating,and the traditional object detection method is difficult to be applied flexibly in the actual scene.Based on this,Open World Object Detection(OWOD)was first proposed by Joseph KJ et al.in 2021,which broke through the closed hypothesis in the mainstream benchmark and applied the object detection model to the open field.The main tasks of this setting are :(1)When there is no clear labeling,the object that has not been learned is identified as ’ unknown ’;(2)When the corresponding label is added,the category of the introduced ’ unknown ’ object is gradually learned,and the category information learned before is not forgotten.Since the setting of OWOD is closer to the actual scene than the existing closed-world static learning settings,it has received extensive attention from researchers.In order to better improve the performance of open world object detection,this paper uses deep learning methods to do the following work :(1)Aiming at the problems of low recall rate of unknown objects and catastrophic forgetting of incremental learning in the existing Open World Object Detection(ORE)method,Adjustable Robust Optimization of ORE based on Graph Feature Pyramid(GARO-ORE)is studied.This method first uses the superpixel image characteristics in Graph-FPN to help the model accurately locate unknown objects and obtain rich semantic information.Then,the robust optimization method is used to comprehensively consider the uncertainty,which greatly ensures that the model avoids forgetting the previously learned category knowledge while learning new categories.Finally,the classification weight initialization method based on knowledge transfer is used to improve the adaptability of the model to new categories.The experimental results on the OWOD dataset show that GARO-ORE achieves better detection results on the recall rate of unknown categories.In the three incremental object detection tasks of 10 + 10,15 + 5 and 19 + 1,the m AP index is increased by 1.38 %,1.42 % and1.44 % respectively.(2)Aiming at the problem that GARO-ORE only considers improving the detection performance of the model and ignores the detection efficiency of the model,a lightweight open world object detection method based on GA-RPN(GAR-LORE)is studied.This method first uses deep separable convolution to replace traditional convolution to optimize the feature extraction network.Then,GA-RPN based on spatial attention mechanism is used to provide a fixed number of candidate regions,and the target candidate regions in the image are extracted lightweightly.Finally,DIOU optimization target regression is used.The experimental results on the OWOD dataset show that GAR-LORE can improve the recall rate of unknown target detection and can lightweight complete the OWOD task.(3)From the perspective of software engineering,an open world object detection system is developed.The open world object detection model proposed in this paper is encapsulated.The B / S architecture model is used to design the system,and the Layui + django + Spring Boot technology stack is used for system development.