This paper presents a new method of an automatic image annotation system that estimates keywords from an image. Typical automatic image annotation systems extract features from an image and recognize keywords. However this method has two problems. One is that it treats features statically. Features should change depending on what keywords are attached so keywords should not be treated equally. Another is that it does not consider the level of keywords. Visual keywords, such as color or texture, can be recognized easily from image features, while high-level semantics such as context are hard to recognize from the features. To solve these problems, our approach is to recognize context by using networked specialist knowledge and to recognize keywords by changing feature values dynamically depending on the context. To evaluate our system, we conducted two experiments of applying it to textile images. As a result, we obtained improved accuracy and confirmed the effectiveness of using networked knowledge.
Automatic image annotation based on deep learning models: a systematic review and future challenges
