Automatic image annotation based on feature fusion and cost-sensitive learning
Zhang Kun'ao, Hao Liu · 2021
In order to solve the problem that it is difficult to recognize small-scale objects in the process of image annotation, an automatic image annotation method based on multi-scale features is proposed. This method combines transfer learning and deep learning, adds a feature fusion layer to the convolutional neural network model, establishes a top-down feature fusion mechanism, and builds a convolutional neural network model that integrates multi-layer convolutional features. The experiment chooses the IAPR TC-12 image annotation data set. Compared with the Tag Prop model, the method in this paper improves the average recall rate by 13%, and improves the average correct rate by 10% compared with the convolutional neural network (CNN-MSE) using the mean square error function. Percentage points. The experimental results show that the convolutional neural network model fused with multi-layer convolutional features can enhance the accuracy of labeling smaller objects.