Research On Vision-based Small Object Detection Method

Object detection is one of the most popular problems in the field of computer vision,and has been widely used in face detection,security monitoring,humancomputer interaction,virtual reality and many other fields.It is of great practical significance to reduce the consumption of human cost through computer vision,so object detection has become the focus of theoretical and application research in recent years.The purpose of object detection is to find the interested object in the image and point out its location.However,the development of object detection is constrained by the problems of illumination change,object occlusion,edge blur,computation cost and so on.This paper summarizes the research status of object detection,and analyzes the existing problems.Then,it introduces the traditional object detection methods and the methods based on deep learning in detail,including the relevant literatures combined with the applications of this paper.On this basis,the detection algorithm based on machine vision and the traffic sign detection algorithm based on Faster R-CNN are proposed,which improves the accuracy and speed of object detection and has high robustness.Finally,this paper summarizes and prospects the research content.In this paper,a fast detection algorithm based on machine vision is proposed.Firstly,the image is preprocessed by Gaussian filtering,gray-scale transformation,edge detection and other methods;secondly,the edge information of parts is extracted by ellipse fitting;then,the edge information of interest is obtained by filtering the edges of different parts;finally,the edge information is extracted and the area is increased or decreased according to the purpose of detection.Experimental results show that this method can effectively solve the problem of illumination and edge information redundancy,and has high accuracy.In this paper,a traffic sign detection algorithm based on Faster R-CNN is proposed.To solve the problem of small object and fuzzy edge,the deconvolution of convolution feature map and the improvement of anchor mechanism are used to obtain more positive samples,and the dislocation sliding window method is used to reduce the calculation cost of the detection system;To solve the problem of too many small objects and the difficulty to label,a detection framework based on weakly supervised learning is proposed,which overcomes the problem of dependence on training samples for object detection.The algorithm has achieved good results in BDCI data set.