Complex Shadow Removal in Pavement Imagery: Leveraging Diffusion Models for Advanced Solutions

Abstract

Data collection is a critical step within road condition monitoring systems. With advancements in data collection methods, vehicle-mounted cameras have become a cost-effective and accessible solution. However, this economic approach often results in images plagued by shadows. These shadows introduce uneven illumination, obscuring crack details and potentially causing shadow boundaries to be misinterpreted as cracks within crack detection algorithms. This study proposes a shadow removal algorithm leveraging conditional diffusion models to eliminate shadows in top-down and oblique-view pavement images. To facilitate this, we introduce the Pavement Image Shadow Triplet dataset (PISTD), based on the Image Shadow Triplet dataset (ISTD). Hyperparameter tuning and training are explored on 256 × 256 and 512 × 512-pixel images to determine the tradeoff between crack detection accuracy and reconstruction precision. Moreover, the two models are compared with state-of-the-art models across peak signal-to-noise ratio (PSNR), root mean squared error (RMSE), and F1 score metrics. Results show that although smaller images result in a 12.3% and 10.3% increase in PSNR, they reduce crack detection accuracy by 23.0% and 17.5% in oblique and top-down views, respectively. Furthermore, in comparison to state-of-the-art approaches, the proposed model achieves a 6% and 38% higher segmentation accuracy on top-down and oblique view images, respectively. An evaluation of the algorithm is performed with real-world shadow images. For this task, the Segment Anything Model (SAM) is utilized in developing the shadow mask pair for shadow removal. Qualitative results on the images demonstrate the effectiveness of the approach.