City Scene Super-Resolution via Geometric Error Minimization

TL;DR

Urban city-scene super-resolution requires preserving geometric structures to support cultural heritage applications. The paper introduces GeoSR, a geometry-aware SR framework built on a UnetSR backbone and augmented with a geometric alignment constraint that leverages Canny edge detection and the Hough transform to produce geometry maps. The losses combine classic geometric error $L_c$, geometric align loss components $L_d$ and $L_p$, into a total objective $L = L_{MSE} + \lambda_d L_d + \lambda_p L_p$, enabling simultaneous improvement in pixel fidelity and geometric consistency. Extensive experiments on Cityscapes and GSV-Cities show GeoSR achieving state-of-the-art PSNR/SSIM, particularly for urban scenes, highlighting the practical impact for cultural heritage preservation, urban planning, and virtual tourism, with code publicly available.

Abstract

Super-resolution techniques are crucial in improving image granularity, particularly in complex urban scenes, where preserving geometric structures is vital for data-informed cultural heritage applications. In this paper, we propose a city scene super-resolution method via geometric error minimization. The geometric-consistent mechanism leverages the Hough Transform to extract regular geometric features in city scenes, enabling the computation of geometric errors between low-resolution and high-resolution images. By minimizing mixed mean square error and geometric align error during the super-resolution process, the proposed method efficiently restores details and geometric regularities. Extensive validations on the SET14, BSD300, Cityscapes and GSV-Cities datasets demonstrate that the proposed method outperforms existing state-of-the-art methods, especially in urban scenes.

Figures (10)

• Figure 1: Visual representations of urban scenes highlighting the geometric features resulting from Hough transform.
• Figure 2: Visual results and geometric features for high-resolution images and SRGAN model reconstructed super-resolution images, with geometric features extraction by Hough transform.
• Figure 3: Compared with previous methods, we introduce a network that constrains geometric loss and pixel loss, effectively preserving the image structure.
• Figure 4: The framework of the GeoSR model derives from the UnetSR lu2022single. The notable modification is to align geometric features between low-resolution and high-resolution images.
• Figure 5: Diagram of the process of geometric feature extraction, involving canny edge convolution and the Hough transform.