Single-Image Shadow Removal Using Deep Learning: A Comprehensive Survey

TL;DR

The paper addresses the challenge of removing shadows from a single image by surveying deep learning approaches that target partial degradation and illumination non-uniformity. It classifies learning strategies (supervised, unsupervised, semi-supervised, zero-shot), discusses network architectures (U-Net, transformer, diffusion), priors (physical illumination, shadow detection), and loss functions, and critically benchmarks methods across standard datasets. The study highlights generalization gaps, computational trade-offs, and emerging directions such as generalized and interactive shadow removal, non-reference metrics, video extension, and larger, higher-resolution datasets, underscoring the practical impact on image quality, downstream vision tasks, and data augmentation. Overall, the survey provides a structured knowledge base and benchmarks to accelerate progress in robust, scalable, and realistic shadow removal research.

Abstract

Shadow removal aims at restoring the image content within shadow regions, pursuing a uniform distribution of illumination that is consistent between shadow and non-shadow regions. {Comparing to other image restoration tasks, there are two unique challenges in shadow removal:} 1) The patterns of shadows are arbitrary, varied, and often have highly complex trace structures, making ``trace-less'' image recovery difficult. 2) The degradation caused by shadows is spatially non-uniform, resulting in inconsistencies in illumination and color between shadow and non-shadow areas. Recent developments in this field are primarily driven by deep learning-based solutions, employing a variety of learning strategies, network architectures, loss functions, and training data. Nevertheless, a thorough and insightful review of deep learning-based shadow removal techniques is still lacking. In this paper, we are the first to provide a comprehensive survey to cover various aspects ranging from technical details to applications. We highlight the major advancements in deep learning-based single-image shadow removal methods, thoroughly review previous research across various categories, and provide insights into the historical progression of these developments. Additionally, we summarize performance comparisons both quantitatively and qualitatively. Beyond the technical aspects of shadow removal methods, we also explore potential future directions for this field.

Figures (11)

• Figure 1: Milestones in deep learning-based single-image shadow removal methods include the exploration of various technologies over time, such as deep CNNs, GANs, RNNs, unrolling, transformers, and diffusion models.
• Figure 2: How a shadow is formed? (a) Shadows are then classified as cast shadows if they belong to the background of the scene or as self shadows if they are part of an occluder object. (b) The cast shadows can be further classified as hard shadows and soft shadows. A crisp edged one (hard shadow) formed by a point light source, a rather more fuzzy one (soft shadow) that is formed by the area light, and otherwise the occluder is very close to the receiver.
• Figure 3: Illustration of common deep learning strategies for single-image shadow removal. Details refer to Section \ref{['sec:learning_strategy']}.
• Figure 4: Illustration of the three major artifacts/challenges in deep learning-based shadow removal, i.e., boundary trace (top), color and illumination inconsistency (middle), and structure distortion (bottom). Results by the representative method for each challenge are shown, i.e., BMNet zhu2022bijective, ShadowDiffusion guo2022shadowdiffusion, and DC-ShadowNet jin2021dc, respectively.
• Figure 5: A statistical analysis of the number of deep learning-based shadow removal methods, according to their learning strategy, network characteristic, mask input, physic model, training dataset, testing dataset, loss function, and evaluation metric.