Self-Supervised Techniques for Intelligent Image Annotation

Abstract

Aiming at the key technical problems such as low correlation between self-supervised learning and image discriminative feature learning, poor adaptation of semantic gap between weakly-supervised information and high-level semantic tasks, and poor generalization ability of incremental updating in the coupling of active learning and migratory learning, this project proposes to carry out the following three main studies. (1) Aiming at the problem of low discriminative ability of self-supervised model learning features, construct a self-supervised learning model with image and target cooperative assistance; (2) Aiming at the problem of poor adaptation of weakly supervised information to the semantic gap among tasks such as target localization, detection and recognition, and segmentation, research on the weakly supervised target discovery framework based on active learning; (3) Aiming at the problem of poor generalization capability of incremental updating of the model, construct a coupling of incremental deep learning models with active learning and transfer learning. coupled incremental deep learning model. We will also construct a multi-point collaborative intelligent annotation system for application demonstration in the fields of smart city and smart medical care.