Deep learning-based method for automatic image annotation

Abstract

Annotating images in order to create tags or meta data for easy searching is often a manual, and thus a time-consuming process. By utilizing a convolutional neural network to perform image classification, and then extracting the features produced by this neural network, one can automate the annotation process, an Autoencoder is then used to compress these feature vectors; this can be considered as a preliminary step which precedes the final task that is to automatically generate basic search tags. These tags can then be used to facilitate the task of image retrieval making the images easier to find, and providing links to other information such as related images and explanatory texts. A very special attention was given to the feature extraction step since it is considered as the cornerstone of automatic image annotation, and without it, the final annotation would not be possible; a statistical analysis and various experiments were conducted on the extracted features in order to determine their characteristics and test out their reliability. To reduce the semantic gap problem, we propose a simple similarity measure method where we assume that each class of tags is represented by a single universal feature vector, this V feature vector then becomes an identifier to that particular class, a new test image’s extracted feature vector will then be compared with this identifier in order to determine whether it belongs to that particular class or not.