Deep-learning-based aesthetic evaluation network for bridge pylon and aesthetics-oriented bridge design

As bridges increasingly serve not only to meet traffic demands but also to fulfill aesthetic expectations, ensuring high aesthetic quality in bridge design has become more important. This paper proposes a deep-learning-based aesthetic evaluation network for the automatic assessment of bridge pylon aesthetics and elaborates on a design method that integrates this network with topology optimization. Firstly, a standardized database and anaesthetic quality evaluation framework specifically for bridge pylons of long-span cable-supported bridges were developed. High-quality bridge pylon data were acquired through a series of image processing methods, followed by an extensive questionnaire to gather aesthetic quality labels for each pylon. Then, different kinds of base models were designed and trained with the labeled dataset, with comparisons made across accuracy and complexity indicators to establish the Pylon Aesthetic Evaluation Network (PAENet). Finally, anaesthetics- oriented bridge pylon design method is proposed which integrates the PAENet with topology optimization, and the design of suspension bridge pylons were given as examples. Through this study, a high-quality database for bridge pylon aesthetics quality evaluation was constructed. Results indicate that the PAENet, using ShuffleNet v2x1 as the base model and enhanced through transfer learning, achieved the highest accuracy in assessing the aesthetic quality of bridge pylons, in alignment with human evaluations. Moreover, the integration of topology optimization and aesthetic evaluation facilitates designs that balance mechanical performance with aesthetic appeal. The proposed method promotes the practical incorporation of aesthetics into bridge design, contributing to the conceptual design of future bridges.