Small Object Detection Algorithm Based On Deep Convolution Neural Network

Object detection is one of the fundamental and important research topics in the field of computer vision.After decades of continuous research,the detection algorithms is becoming more and more robust.What is even more exciting is that in recent years,as deep learning technology has made major breakthroughs in the field of computer vision,the object detection framework based on deep learning has become the mainstream and gradually derived based on candidate regions and regression-based deep learning.There are two main branches of the object detection algorithm.Candidate region detection methods based on the R-CNN series greatly improve the accuracy of object detection tasks.However,detection methods based on candidate regions have problems such as cumbersome procedures and slow detection speeds.Therefore,the regression-based detection method directly returns the object on the input image,eliminating the process of extracting the candidate region,making the detection speed achieve real-time results,but the detection accuracy is slightly reduced.The detection precision and detection speed always restrict the practicality of the target detection algorithm.Moreover,the above two types of detection frameworks have the problem of easy-missing detection in complex scenes,especially the detection of small objects.This Thesis summarizes the domestic and foreign research progress,basic research framework and latest research results of the object detection algorithm.It focuses on the application of deep learning techniques in object detection and analyzes possible problems with detectors.Faced with the more difficult problem of small target detection,this Thesis proposes a new object detection algorithm based on the fast target detection based on candidate regions to solve the problem of small target detection in traffic scenes.The main tasks of this Thesis are as follows:(1)Faster R-CNN is one of the typical representatives of the deep learning object detection algorithm.The mAP on the VOC 2012 test set reached 70.4%.However,it is easy to miss the problem of small objects.To solve this problem,this Thesis fuses multi-layer convolution maps,and uses low-level convolution features to locate small objects,to obtain more accurate positioning information,and to use high-level convolution features to medium and large objects.fusion of multi-layer convolution features,generate rich and robust features,is conducive to the detection of small objects.An online hard example mining algorithm is used to update the model parameters,which further accelerates the convergence of the model,solves the problem of imbalance in the number of training samples,and improves the generalization ability of the model.This can effectively solve the problem of missed detection of small targets while ensuring the overall detection effect.(2)The improvement of accuracy is often at the expense of reduced detection speed,and the accuracy of real-time detection algorithms is not high.For the object detection algorithms requires real-time and high-precision requirements,this Thesis proposes a object detection algorithm that takes into account the detection speed and accuracy.By constructing a right-to-left,top-to-down feature pyramid,the detectors at different scales correspond to detectors of different scales.At the same time,features rich in detail information and rich semantic information are generated,which is favorable for obtaining more accurate positioning information,thereby improving detection.The accuracy of the algorithm,especially to improve the detection ability of small objects.In addition,in order to achieve real-time detection speed,this Thesis constructs target priors on the fusion feature map,greatly reduces the sample search space,and effectively improves the object detection speed.(3)The algorithm proposed in this Thesis is validated on the KITTI data set of autopilot scenario,PASCAL VOC 2007 data set and PASCAL VOC 2012 data set,and compared with the current mainstream detection algorithm.In the complex scene,the algorithm of this Thesis effectively improves the ability to detect small objects and improves the detection accuracy as a whole while ensuring the overall detection speed.In addition,this algorithm has a certain ability to deal with partial occlusion.