Research On Small Object Detection Based On Deep Learning

Object detection is one of the most important research issues in the field of computer vision.After decades of development,especially the rise of deep learning in recent years,the effect of object detection algorithms has been continuously improved,and its robustness has been stronger.The main research direction of object detection is based on deep learning,and it includes two types: object detection algorithms based on candidate region and object detection algorithms based on regression.The object detection algorithms based on candidate region which are represented by Faster R-CNN has a great improvement in the accuracy of detection,but its detection speed is slow.The object detection based on regression which are represented by SSD improve the detection speed,but the accuracy of detection decreases.Although the research of object detection algorithm has made great progress in recent years,most of the object detection algorithms focus on the object with ordinary size,the accuracy of detection is not satisfactory for special application scenarios such as small objects.Therefore,this paper has carried out the research on the problem of small object detection.This paper summarizes the research process and research status of object detection algorithm at home and abroad.In order to solve the problem of object detection in the small object scenarios,this paper analyzes and validates several effective methods to solve the small object detection problems,and combines these methods to propose a small object detection algorithm based on the regression-based object detection framework.The object detection algorithm has achieved great results on the Wider Face dataset.The main research work of this paper is as follows.Based on the three research points of context,image up-sampling and the design of temple,this paper designs image classification experiments to verify the effectiveness of feature fusion,multi-layer feature prediction and image pyramid for the small object detection.In order to adapt to the size of object in the dataset of the small objects,this paper designs 25 templates based on the distribution of the object size of the Wider Face dataset using k-means clustering method.Based on feature fusion,multi-layer feature prediction and image pyramid,this paper proposes a object detection algorithm for small object problem based on the regression-based object detection framework,and the Faster R-CNN,SSD,MTCNN algorithms are compared from the detection accuracy and detection speed on the Wider Face dataset.According to the experimental results,the object detection algorithm proposed in this paper is much better than other classical object detection algorithms in the small object dataset.Therefore,the object detection algorithm of this paper is effective for the small object detection problems.