Research on Aerial Image Road Extraction Method Based on Improved SRCNN

Abstract

Aerial image road extraction is a key task in intelligent transportation and urban planning. In response to the limited feature extraction ability, insufficient multi-scale adaptability, and weak road structure preservation ability of existing lightweight super-resolution networks (SRCNN) in this task, this paper proposes an enhanced SRCNN model (ISRE) that integrates structure perception. This model integrates an efficient channel attention module in the SRCNN framework to focus on road features, designs a multi-scale feature fusion module to capture road shapes at different scales, and constructs an adaptive structural perception loss function that integrates pixel, gradient, and region constraints to optimize topology continuity. Experiments using the Massachusetts road dataset demonstrate that ISRE enhances its F1 score by 1.2% (relative) and topological accuracy by 11.2% in comparison to the original SRCNN, while preserving model efficiency, and its overall performance surpasses that of more sophisticated models like U-Net. This study provides an effective solution for efficient road extraction in resource constrained scenarios, and will explore directions such as Transformer enhancement and edge device deployment in the future.