Semantic and Instance Segmentation in Coastal Urban Spatial Perception: A Multi-Task Learning Framework with an Attention Mechanism

With the continuous acceleration of urbanization, urban planning and design require more in-depth research and development. Street view images can express rich urban features and guide residents’ emotions toward a city, thereby providing the most intuitive reflection of their perception of the city’s spatial quality. However, current researchers mainly conduct research on urban spatial quality through subjective experiential judgment, which includes problems such as a high cost and a low judgment accuracy. In response to these problems, this study proposes a multi-task learning urban spatial attribute perception model that integrates an attention mechanism. Via this model, the existing attributes of urban street scenes are analyzed. Then, the model is improved by introducing semantic segmentation and instance segmentation to identify and match the qualities of the urban space. The experimental results show that the multi-task learning urban spatial attribute perception model with an integrated attention mechanism has prediction accuracies of 79.54%, 78.62%, 79.68%, 77.42%, 78.45%, and 76.98% for the urban spatial attributes of beauty, boredom, depression, liveliness, safety, and richness, respectively. The accuracy of the multi-task learning urban spatial scene feature image segmentation model with an integrated attention mechanism is 95.4, 94.8, 96.2, 92.1, and 96.7 for roads, walls, sky, vehicles, and buildings, respectively. The multi-task learning urban spatial scene feature image segmentation model with an integrated attention mechanism has a higher recognition accuracy for urban spatial buildings than other models. These research results indicate the model’s effectiveness in matching urban spatial quality with public perception.