| Image annotation is a repetitive and tedious process.When the number of pictures is relatively small,such as hundreds or thousands of pictures,manual labeling can be used.On the one hand,it can ensure the accuracy of labeling,and on the other hand,it will not take too much time.However,when the number of pictures reaches tens of thousands or even hundreds of thousands,it is very difficult to manually label these pictures one by one,and researchers need to spend a lot of time and energy on repetitive labeling tasks.Therefore,many research teams hand over the labeling work to outsourcing companies,but this will not only generate additional economic costs,but also cause data leakage to some extent,especially when the data set is of a confidential nature.In order to reduce the time cost and economic cost of researchers on labeling tasks,while ensuring data security to the greatest extent,this paper designs a locally deployable image automatic labeling system based on Adversarial FasterRCNN.The main design and implementation work of this paper are as follows:(1)The image automatic annotation system designed in this paper is based on the Adversarial Faster RCNN algorithm.In order to improve the ability of Faster RCNN to detect occluded objects,this paper builds Adversarial Faster RCNN by fusing data enhancement network on the basis of Faster RCNN and uses data enhancement network to generate adversarial training samples that are difficult for Faster RCNN to distinguish.Faster RCNN’s ability to distinguish occluded objects.After testing,the mAP of the Adversarial Faster RCNN designed in this paper can reach 74.6%on the PASCAL VOC 2007 test set,which is slightly higher than the 73.2%of Faster RCNN.(2)The image automatic labeling system designed in this paper can use the pre-trained model to automatically label the data set.The construction of the image automatic labeling system is based on the strategy of separating front-end,back-end and data.Users can read local image data through the front-end interactive module,and then upload the data to the back-end service module,which calls the algorithm to label the data,and store the annotation data in a separate data server to ensure the consistency of system data.When the labeling fails or the labeling information is incomplete,the user can add,delete,modify,query the labeling results through the manual labeling function.(3)This paper tests the system’s front-end interaction module,backend service module,data management module,system scalability,annotation speed and annotation quality.The test results verify the effectiveness and reliability of the system’s functions.This system can achieve the expected design goals,and when the same labeling task is completed in the specified test environment,this system can save more than 93%of the time compared with manual labeling and can achieve 68.1%of the quality of manual labeling. |
