Domain-Specific Image Classification Using Ensemble Learning Utilizing Open-Domain Knowledge

Image classification is a fundamental research problem with a lot of applications in computer vision. The labels in most existing research in image classification are concepts from natural objects or scenes (e.g., cat, dog, sky, mug). Most image classification tasks conduct training and testing using pre-labeled datasets due to its supervised fashion. However, in the real-world scenarios, images need to be classified into the labels from more abstract concepts (e.g., immodest images, offensive images). These concepts usually come from a specific domain and are defined by some guidelines. We call this type of problem domain-specific image classification. Lack of labeled images for training, especially for domain-specific images, is a common challenge in image classification. In this paper, we propose a novel framework to address domain-specific image classification with insufficient labeled data for training. Since a label from an abstract concept usually includes multiple specific concepts or labels, we decompose each tag with the abstract concept into multiple labels with each label corresponding to a specific concept using domain knowledge. Then we use each specific concept as a base learner, and combine the multiple base learners to form an ensemble learning. For each base learner, we utilize open-domain knowledge (e.g., search engines) to obtain labeled data to train each base learner with deep neural networks. Finally, an ensemble learner is trained using another neural network by incorporating these base learners. Using immodest images as an example of an abstract concept, experiments have been conducted to show that the proposed neural-network-based ensemble learning performs better than traditional ensemble learning. Also, the combined strong leaner with all base learners performs better than a subset of base learners. In addition, this ensemble learning framework reduces the dependency of in-house labeled data by utilizing public-domain knowledge and still reaches the comparable results.