Abstract: Cross-domain discrepancy frequently hinders deep neural networks to generalize to different datasets. In order to improve the robustness of shadow detection, a novel unsupervised domain adaptive shadow detection framework was proposed. Firstly, in order to reduce the data bias between different domains, a multi-level domain adaptive model was introduced to align the feature distributions of source domain and target domain from low level to high level. Secondly, to improve the model ability of soft shadow detection, a boundary-driven adversarial branch was proposed to guarantee the structured shadow boundary was also able to be obtained by the model on the target dataset. Thirdly, the entropy adversarial branch was combined to further suppress the high uncertainty at shadow boundary of the prediction result, so as to obtain an accurate and smooth shadow mask. Compared with the existing deep learning-based shadow detection methods, the proposed method has the Balance Error Rate (BER) averagely reduced by 10.5% and 18.75% on ISTD dataset and SBU dataset respectively. The experimental results demonstrate that the shadow detection results of the proposed algorithm have better boundary structure.

Key words: unsupervised domain adaptation, deep learning, image processing, shadow detection, information entropy

摘要： 跨域差异常常会阻碍深度神经网络的泛化，使其不能适应不同的数据集，为了提高模型阴影检测的鲁棒性，提出了一种新颖的无监督域适应阴影检测框架。首先，为了缩小域间的数据偏差，采用分层域适应策略校准源域和目标域间从低层到高层的特征分布；其次，为了加强模型软阴影的检测能力，提出边界对抗分支以确保模型在目标数据集上同样可以得到结构化的阴影边界；然后，结合熵对抗分支进一步抑制预测结果中边界处的高不确定性，从而得到边界平滑、准确的阴影掩膜。与已有深度学习检测方法相比，所提方法在客观数据集ISTD、SBU上的平衡误差率（BER）分别降低了10.5%、18.75%。实验结果表明所提方法的阴影检测结果具有更好的边缘结构性。

关键词: 非监督域适应, 深度学习, 图像处理, 阴影检测, 信息熵