Making a New City Image

This paper explores the application of computer vision and machine learning to streetlevel imagery of cities, reevaluating past theory linking urban form to human perception. This paper further proposes a new method for design based on the resulting model, where a designer can identify areas of a city tied to certain perceptual qualities and generate speculative street scenes optimized for their predicted saliency on labels of human experience. This work extends Kevin Lynch’s Image of the City with deep learning: training an image classification model to recognize Lynch’s five elements of the city image, using Lynch’s original photographs and diagrams of Boston to construct labeled training data alongside new imagery of the same locations. This new city image revitalizes past attempts to quantify the human perception of urban form and improve urban design. A designer can search and map the data set to understand spatial opportunities and predict the quality of imagined designs through a dynamic process of collage, model inference, and adaptation. Within a larger practice of design, this work suggests that the curation of archival records, computer science techniques, and theoretical principles of urbanism might be integrated into a single craft. With a new city image, designers might “see” at the scale of the city, as well as focus on the texture, color, and details of urban life.