Design And Implementation Of Image Category Annotation System Based On Noisy Label Correction

Deep learning utilizes multi-layer nonlinear neural network to extract and represent information,and has achieved remarkable results in the field of computer vision.The success of deep learning relies on a large amount of high-quality annotated data for supervised learning,but the labor cost of a large amount of annotated data is huge.And marked by the package or web crawler lack of data on the reliability of the data quality guarantee,the introduction of these methods will inevitably label noise,and negative influence on the image classification task,therefore,to build a fully functional and can correct label noise image data labeling system is important.The current image category annotation system relies on the data crowdsourcing platform.By disassembling a large number of data annotation tasks into several sub-tasks and then assigning them to manual annotation,the annotation platform will recycle these annotation tasks to form a complete data set after annotation.This kind of system faces the following challenges:1)the sub-task verification is difficult,and the audit cost brought by several sub-tasks is huge;2)There is noise in image annotation,and the large-scale image data set obtained through the crowdsourcing system will inevitably introduce error annotation.So,in view of the existing platform in the first paragraph is insufficient,we constructed based on noise images of the correct category label labeling system,the system integrates a automatic correcting module for automatic reading labels image data stored in the database and automatically correct potential image label,to complete the revised new tag will be written back to the database.Aimed at the existing platform second is insufficient,we innovative was proposed based on the label of the noise figure neural network inference model,the model through the construction of a dual graph structure to capture the structure of the relationship between two different levels of the label(including instances and distribution relationship),provide the supervision of the clean signal for depth of neural network,so as to offset label noise.His research results were published in the Proceedings of IEEE/CVF Conference on Computer Vision and Pattern Recognition(CVPR)2021.In order to test the completeness and robustness of the system,we constructed a large number of functional test cases,which proved that our system was superior to other image labeling systems.