Abstract
Removing rainy artefacts (e.g., rainstreaks and raindrops) can benefit advanced computer vision systems, such as intelligent surveillance systems and self-driving cars. Data-driven models produce more promising results than model-based approaches. However, by evaluating representative deep-learning based deraining models on benchmark datasets, two unfavorable results are obtained: (1) they cannot handle various rainy conditions (i.e., rains of various directions and densities) equally well, and (2) the models trained on synthetic images cannot generalize well to process real rainy images.
This Ph.D. thesis has made progress towards solving both issues. A deraining network dissection tool (DNDT) was firstly designed to analyze the behavior of filters in deraining models. Task-specific issues are then identified and solved by formulating a series of novel deraining models: (1) it was found that the imbalanced distribution (between rains and backgrounds) results in biased representation, and then a context-enhanced representation learning and deraining network (CERLD-Net) was proposed to learn an explainable and controlled representation; (2) it was found that the representations learned in existing models suffer from the loss of features depicting the background, and then an entangled representation learning network (ERL-Net) was proposed to learn more complete features; (3) it was found that some negative factors are learned in the latent space of ERL-Net, and then a disentangled representation learning and enhancement network (DRLE-Net) was proposed to embed a rainy image into two separated spaces such that only the positive factors are selected for deraining; and (4) by dissecting DRLE-Net, the significance of using rain distribution map (RDM) as prior is verified, motivating the formulation of a novel cascaded attention guidance network (CAG-Net) for better exploiting RDM in a hierarchical manner.
The issues of the coarse-to-fine deraining models are also identified and solved by proposing a novel attentive feature refinement network (AFR-Net) to remove heavy rainy artefacts.
In summary, this thesis investigates and solves the problems of existing deraining models by proposing novel networks, which are designed towards learning unbiased and explainable representations for reconstructing clean images without over-/under- deraining artefacts. Extensive experiments demonstrate that the proposed methods outperform the state-of-the-arts by large margins.