The Research Of Weakly Supervised Object Detection Based On Attention Mechanism

The task of Object Detection is one of the basic problems in the field of Com-puter Vision,and the object detection algorithm based on supervised learning is the mainstream of the current research field,which leads to the necessary premise that a high-precision tagged data set can still obtain good performance in most cases.How-ever,in many real production and life scenes,the high cost of image acquisition and the difficulty of tagging by proficients lead to the acquisition of high-quality image tagging is very time-consuming and laborious.With the continuous development of deep learn-ing technology and theory,how to train high-quality object detection model through low-quality training data has high research significance and practical value.To solve this problem,in this article,it proposes a weak supervised object detection algorithm based on attention mechanism.The ability of using attention mechanism to filter information is used to strengthen the attribute of object location of Convolutional Neural Network(CNN).And based on this,a model of rough object detection in training place under weak supervision data set is realized.In this ariticle,some problems of attention mechanism are also studied and experimented.By adding attention map merging algorithm and result refinement algorithm to the whole algorithm,the overall effect of weak supervised object detection model is further improved.The main research contents and results are as follows:1.By imitating the attention mechanism of human beings in understanding things and applying it to the task of weak supervised object detection through algorithm,the goal of training high accuracy object detection model with only image level tagged data is realized.2.The characteristics of attention mechanism will lead that it tends to local optimal solution and focus on the local information of the target to be detected,which also leads to the output of attention mechanism can not be directly used in the actual object detection task.In order to solve this problem,this ariticle designs the attention map merging algorithm,which trains many different attention models repeatedly by similar data enhancement,and obtains more valuable object detection results by synthesizing the results of these models.3.In order to further optimize the results of object detection,combined with the traditional computer vision edge detection algorithm,this ariticle designs a coarse result refinement algorithm based on edge detection for the classic object detection task of single object detection,and conducts relevant experiments on Pascal VOC data set.It proves that,after combining edge detection algorithm,the weak supervised object detection algorithm designed in this paper can achieve the same effect as the strong supervised object detection algorithm,and also proves the generalization ability of the algorithm in medical area.4.Aiming at the natural scene text detection task,we design and use a result refinement algorithm based on character detection algorithm,combining with statis-tical method.We also carried out experiments on MSRA-TD500,ICDAR2013 and ICDAR2015,and achieved 81%,82%and 59%accuracy respectively.