Research On Low Resolution Small Object Detection Technology

Region Proposal Strategy With the entry of human society into the 21 st century,the accumulation of various massive data and the substantial increase in computing power have led to the rapid development of deep learning technology.Deep learning has made breakthroughs in object recognition,detection,image generation and many other image fields.Object detection technology has been widely used in industrial production,public transportation,medical and health fields,and promotes the intelligent process of production and life.Small object detection at low resolution has infinite potential in areas such as urban transportation,industrial production,and mass entertainment.Most of the existing object detection technologies are focused on large size objects at high resolution,and their detection technology is mature with excellent detection performance.However,small object detection at low resolution is still a major challenge in the field of object detection due to object size and the poor resolution.This thesis is based on existing deep learning methods,for solving the problem of small object detection at low resolution,research on related technologies has been carried out.The main work of thesis includes two aspects: data set construction and expansion,detection algorithm construction and improvement.On the one hand,this thesis analyzes the existing object detection datasets and finds that the existing datasets are mostly for large objects at high-resolution detection tasks,which cannot meet the research needs of this thesis.In this thesis,some low-resolution small object images are extracted from the existing public dataset,and ten small object datasets SmallObjects-10,which meet the research needs of this paper,are constructed by hand-labeling.In addition,artificial synthetic methods are used to expand data set.On the other hand,based on the SmallObjects-10 dataset,three different deep neural network algorithms are designed to detect and identify by introducing image context information and changing the region proposal strategy.The effectiveness of these methods is verified by experiments and the speed and accuracy of different algorithms are compared to analyze their advantages and disadvantages.In summary,the research on the detection of small objects at low resolution by using the deep learning method fully demonstrates the feasibility and advantages of deep learning in small objects at low resolution.In addition,the accuracy of the detection method is improved by the method of data expansion.