A Multi-model Learning Method for Geographical Emotion Detection Based on Street View Images
As an important way to sense the urban living environment,vision has been playing a vital role in dis-covering the urban physical environment.As a type of ubiquitous geospatial data,the street view images describe the urban environment from the perspective of visual perception,which makes it popular in typical place classifica-tion and spatial characteristic analysis researches.In previous researches,the ability of street view images has been proved in measuring human perceptions,but the rich contexts in images have been hardly fully utilized.A learning method using the semantic segmentation and natural language processing is proposed for measuring human emotion perceptions.By embedding geographical entities into the word vector and concatenating with visual feature vector,fusion features show better performance in the classification task.The results indicate that the proposed method learns meaningful representations comparing to existing methods,which could significantly improve performance in the emotion prediction task.And the information from street images can be fully extracted,which contributes to a better understanding of the underlying urban environments.
street view imagesdeep learningsemantic segmentationgeospatial big datanatural language pro-cessing