Urban Street Quality Measurement in Central City by Combining Street View Images and Deep Learning

With the current transformation of urban planning and design refinement, the study of street spatial quality characteristics under the humanistic scale has received more and more attention. A comprehensive use of quantitative evaluation methods for street spatial quality measurement is needed to establish a multi-dimensional refinement and scientific analysis framework under the humanistic perspective. In this study, we introduce panoramic images of Shenyang central city, use deep convolutional neural network model to give six types of emotion scores in streetscape, extract the characteristics of visual state and pixel elements of streetscape through image scene segmentation algorithm, use linear regression to explore the influence of standardised factors on subjective perception scores, and analyse how the elements of perception scores can improve the quality of street space. The results show that: ① Considering the frequency and conditions of the streets used using street perception scores and distribution analyses provides an accessibility that is consistent with a humanistic perspective in everyday life. ② Perceptions of high affluence and safety are widely distributed in the centre of the city, suggesting that the overall spatial quality of the area is good. Lively has high levels in the West Midlands and Boring dominates on the city’s main roads. Trunk roads contain a significant amount of traffic and therefore consideration should be given to increasing the greening rate and adding planted streetscape elements to alleviate boredom. ③ Based on the distribution of the six categories of perceived correlations mapped to the top 25 percent of high perception scores it was found that the correlations between Safe, Lively and Affluent were all highly positive on streets with high perception scores. The spatial distributions were very similar, occurring mainly in the West Midlands and more dispersed in the North. The boredom and frustration dimensions have an overall fragmented distribution, with similarities between the two. Streets with these types of spatial perception scores need to be further improved to effectively enhance the positive experiences of residents. ④ A review of street perception scores and mapping distributions through linear regression analyses revealed that beauty had a significant positive effect with both sky and plants, and a significant negative effect on frustration. In positive spatial perception, plants had a significant effect on the beauty dimension, while other visual elements had a significant negative effect. The spatial comfort atmosphere can be promoted by the percentage of road, street furniture, sky and green elements, which in turn improves the spatial quality of the street. This study provides basic information about urban spatial perception patterns. Considering the distribution of affective perceptions and street quality measures in deep learning can provide methodological support for finer urban planning research, with a view to providing references and suggestions for the research and practice of urban street spatial quality.