An improved BK sub-triangle product approach for scene classification

Scene classification is a popular research topic in computer vision, and has received much attention in the recent past. Conventionally, scene classes are considered to be mutually exclusive. However, in real-world scenarios a scene image may belong to multiple classes, depending upon different perceptions of the masses. In this paper, we propose an improved Bandler and Kohout’s sub-triangle product (BK subproduct) approach to address this issue. Instead of using the original BK subproduct solely, we introduce a combination of inference structures. The advantages are three-fold. Firstly, using the BK subproduct as an inference engine, we are able to attain the relationships between image data sets and scene classes that are not directly associated. Secondly, our approach is able to model non-mutually exclusive data, as opposed to conventional solutions. Finally, our classification result is not binary. Instead, we can classify each scene image as belonging to multiple distinct scene classes. Experimental results on public datasets demonstrate the effectiveness of the proposed method.