Data equilibrium based automatic image annotation by fusing deep model and semantic propagation

Highlights：

• Image annotation task is regarded as a multi-label classification problem to be solved using a stacked auto-encoder (SAE).

• A robust balanced and stacked auto-encoder (RBSAE) is proposed to improve the effectiveness of training biased datasets and the stability of model training.

• The method of constructing an ideal local equilibrium dataset is proposed, which may become a promising approach for annotating image.

• A semantic propagation algorithm based on the local equilibrium dataset (LDE-SP) is introduced to promote the annotation accuracy of middle- and low-frequency tags.

• Non-linear and linear optimization methods are used to train the SAE models for solving the image annotation problem.

• The proposed annotation method that discriminates the high- and low-frequency attributes of images (ADA) can effectively improve the non-zero recall value (N+).