Distraction-Aware Edge Enhancement for Shadow Detection in Remote Sensing Images

Currently, the complex features of remote sensing images complicate shadow detection, such as blurred shadow edges, a large number of fragmented shadows, and varying shadow colors. To address these challenges, a distraction-aware edge enhancement for shadow detection network (DESDNet) is proposed. It utilizes distraction-aware shadow (DS) modules to reduce false positives (FPs) and false negatives (FNs), while employing edge enhancement modules to improve shadow edge detection. In this network, the shallowest convolutional layers capture detailed shadow edge features, including many nonshadow background details, while the deepest convolutional layers, with larger receptive fields, effectively suppressing nonshadow pixels. The edge module integrates the two features to mitigates issues, such as edge blurring, fragmentation, and color inconsistency in shadows. The proposed network has been validated for feasibility on remote sensing image datasets, on which our experiments achieve the accuracies of 98.48% on the ISTD dataset and 96.70% on the SBU dataset, outperforming typical networks recently. The code is available from
https://github.com/sfs0/DESDNet/tree/main
.