Research on Street Space Quality Combined with Attention Multi-Task Deep Learning

Abstract

In the context of advocating human-oriented urbanism, in order to more accurately evaluate the urban scene. This paper uses the sentiment evaluation of millions of street view images in the Place Pulse 2.0 data set to design a multi-task deep learning network (Mtl-attn-vgg) fused with attention to analyze the urban spatial quality attributes. Multi-task learning is used to simultaneously learn different attributes and the attention mechanism to focus on the characteristics of the object to improve the effect of the convolutional layer on the attribute feature extraction, combined with deep relative attribute learning, use the structural sparsity of the feature matrix to compare the attribute features In addition, classification loss and ranking loss are introduced to constrain the parameters of the deep learning network. Finally, six attributes of street view are beautiful, boring, depressed, lively, safe, and rich. The results show that the Mtl-attn-vgg model improves the feature extraction effect and improves the accuracy of attribute sorting to 78.48%, which is more conducive to urban designers and planners to understand the attribute characteristics of the street scene environment.