The visual quality of streets: A human-centred continuous measurement based on machine learning algorithms and street view images

This study proposes a workable approach for quantitatively measuring the perceptual-based visual quality of streets, which has often relied on subjective impressions or feelings. With the help of recently emerged street view images and machine learning algorithms, an evaluation model has been trained to assess the perceived visual quality with accuracy similar to that of experienced urban designers, to provide full coverage and detailed results for a citywide area. The town centre of Shanghai was selected for the site. Around 140,000 screenshots from Baidu Street View were processed and a machine learning algorithm, SegNet, was applied to intelligently extract the pixels representing key elements affecting the visual quality of streets, including the building frontage, greenery, sky view, pedestrian space, motorisation, and diversity. A Java-based program was then produced to automatically collect the preferences of experienced urban designers on representative sample images. Another machine learning algorithm, i.e. an artificial neural network, was used to train an evaluation model to achieve a citywide, high-resolution evaluation of the visual quality of the streets. Further validation through different approaches shows this evaluation model obtains a satisfactory accuracy. The results from the artificial neural network also help to explore the high or low effects of various key elements on visual quality. In short, this study contributes to the development of human-centred planning and design by providing continuous measurements of an 'unmeasurable' quality across large-scale areas. Meanwhile, insights on the perceptual-based visual quality and detailed mapping of various key elements in streets can assist in more efficient street renewal by providing accurate design guidance.