Assessing visual similarity of neighbourhoods with street view images and deep learning techniques

Despite the wide availability of street-view data and advanced computational techniques, the topic of perceived visual similarity in urban design has received little attention. The impact of visual sameness on the loss of urban identity and its effect on individuals' health has been widely debated. However, empirical evidence to support these arguments has been limited. This study proposes a set of tools to measure similarity in urban neighbourhoods. It utilizes Street view images and DL models such as semantic segmentation and generative inpainting for image enhancement and refinement. It further employs the LPIPS, a DL-based metric that computes image-based perceptual similarity.