Estimating aesthetic services of road landscapes through predicting people's attention: A computer vision approach.

The aesthetic services of road landscapes provide recreational opportunities for the road environment, thereby supporting the designation, planning and design of scenic roads. Computer vision presents a methodology to investigate landscape aesthetic services by offering pixel-level tools to identify and analyse people's aesthetic attention. These tools can help overcome some of the limitations of examining attention through eye-tracking experiments. In this study, we constructed a dataset by collecting image data of road landscapes in Southwest China and creating aesthetic labels through public ratings. We employed a two-step deep transfer learning to train an aesthetic prediction model. The resultant model presented an accuracy of 0.88 in identifying landscapes with notable aesthetic features. Then we leveraged a class activation mapping to elucidate the model's aesthetic attention in the image samples. To interpret the visual features of aesthetic attention, we adopted image segmentation, colour extraction, depth estimation and edge detection to analyse the elements, colours, deepness and complexity of the attention areas in landscapes. Our results demonstrated the different patterns between positive and negative aesthetic attention. Negative attention is focused on unattractive objects, gravitating towards nearby artificial objects with dull colours and basic outlines. In contrast, positive attention displays a preference for distant, brightly coloured natural objects with complex shapes. Its pattern involves more than just the search for attractive objects, as it also includes a general focus on the landscapes around the road end and roadsides. The proposed approach can be used to estimate the aesthetic services of road landscapes, and the empirical findings offer implications for the planning and design of scenic roads.