Design and Development of Scene Recognition and Classification Model Based on Human Pre-attentive Visual Attention

Recent works on scene classification still utilize the advantages of generic feature of Convolutional Neural Network while applying object-ontology technique that generates limited amount of object regions. Human can successfully recognize and classify scene effortlessly within short period of time. By utilizing this idea, we present a novel approach of scene classification model that built based on human pre-attentive visual attention. We firstly utilize saliency model to generate a set of high-quality regions that potentially contain salient objects. Then we apply a pre-trained Convolutional Neural Network model on these regions to extract deep features. Extracted features of every region are then concatenated to a final features vector and feed into one-vs-all linear Support Vector Machines. We evaluate our model on MIT Indoor 67 dataset. The result proved that saliency model used in this work is capable to generate high-quality informative salient regions that lead to good classification output. Our model achieves a better average accuracy rate than a standard approach that classifies as one whole image.