| Summary: | 碩士 === 國立成功大學 === 製造資訊與系統研究所碩博士班 === 101 === “Images” as writings, are created, taken or edited by authors, which might carry the intent of the authors. In previous researches, “image annotation” method was used to indicate the concepts of image. However, object labels which are the outputs of “image annotation” can’t describe the meanings completely. This study proposes a method for image conceptualization by referencing the image realization process of humans. First of all, objects are annotated and scene is classified according to the relations between objects. Then an event of the image is inferred based on the objects and the scene identified in the previous step. Finally, the concept of the image is described using a simple sentence. Owing to the fact that most of images are living photos, this study focused on living photos for experiments of image conceptualization. The experiment showed that the accuracy of sentences for image concept description, which were evaluated by humans, is up to 87.3%. It also proved the applicability of the proposed image conceptualization method.
|
|---|
