Applications of Modelling and Simulation (Jun 2024)
Building Detection based on Faster RCNN with Distributional Soft Actor-Critic with Three Refinements
Abstract
This research presents a comprehensive framework for building detection in high-resolution images, integrating advanced techniques from computer vision and reinforcement learning. The methodology employs the Faster Region-based Convolutional Neural Network (RCNN) architecture for efficient feature extraction and region proposal generation, enhancing the accuracy of building detection. A novel Fine Region Proposal Network (FRPN) adapts region proposals based on image characteristics, dynamically adjusting candidate regions for improved efficiency. The study introduces three refinements to the Distributional Soft Actor Critic (DSAC-T) algorithm, addressing stability and sensitivity concerns. These enhancements involve fine-tuning the critic gradient, incorporating twin value distribution learning, and introducing a variance-based mechanism for return clipping the target. Rigorous assessments on demanding datasets, such as the Massachusetts and WHU building dataset, provide compelling evidence of the efficacy of the proposed framework. The proposed approach demonstrates superior performance in building detection, achieving an average precision of 69.48% and an average recall of 84.29% on the Massachusetts dataset, and an average precision of 65.82% and an average recall of 81.52% on the WHU dataset. Thus, the research contributes to the field by providing a robust solution for building detection, leveraging state-of-the-art techniques for improved performance in diverse urban and suburban environments.
