Research On Small Object Detection Framework Based On Attention Mechanism

Small object detection is a task for detecting low-sized pixel objects in images.It has always been of high research value and has broad application prospects in UAV reconnaissance,satellite remote sensing,autonomous driving,and industrial intelligent manufacturing.During the research process,this thesis has noticed that there is currently a valuable research direction in the field of deep learning: attention mechanisms.For small object detection and attention mechanisms,small object detection is a subdivision of object detection.Nowadays,with the highly developed object detection,small object detection still has a lot of space for progress;However,attention mechanism,as an efficient optimization method for deep learning in recent years,has not been successfully integrated with small object detection deeply.Therefore,this thesis has proposed a small object detection framework based on attention mechanism,specifically integrating attention mechanism into small object detection,solving the problem of small objects being difficult to identify due to low pixel size,and improving the possibility of wide application of small object detection in multiple scenes.The main research contents of this thesis are as follows:Firstly,this thesis has proposed an Attention Oriented Deep Slices(AODS)Detection Framework through in-depth analysis of common detection strategies,main problems and common solutions of small object detection,which also combined with the analysis of the advantages and disadvantages of attention mechanisms.This framework proposes a new definition for small object area-Pixels Quantities Ratio(PQR)definition,which defines whether the area is worth to be detected by considering all object pixels in the area.Compared with the traditional definition method,it uses more pixels for collaborative definition.At the same time,AODS framework will adopt different object detection strategies according to the classification results defined by PQR,such as deep slice detection of small object areas,to ensure the effect and efficiency of small object detection.Secondly,under the PQR definition,in view of the characteristics of objects and backgrounds with different sizes that need to be considered for small object classification,this thesis has designed a Mixed and Dilated Pyramid Attention(MDPA)network that can quickly classify small object regions.This network can expand the kind of the multiscale receptive field,optimize the amount of parameters,and set up a multi branch attention mechanism,it achieves a fast and effective detection of whether there are small objects in the target area through comprehensive analysis of all pixels in the target area.Finally,this thesis has analyzed the advantages and disadvantages of AODS framework,PQR definition,and MDPA network compared with existing methods through experiments,summarized and predicted the potential directions of attention based small object detection algorithms in future research.