InNet: Learning to Detect Shadows with Injection Network

Abstract

Shadows bring great challenges but also play essential roles in image understanding. Most recent shadow detection methods are based on patches, then further reasoning method is required for the obtaining of a completed shadow detection result for an image. In this paper, an injection network is proposed to detect shadow regions for the whole image directly. In order to maintain as many as details, the skip structure is applied to directly inject the details from convolutional layers to de-convolutional layers. Meanwhile, a weighted loss function is proposed for the network training. With this adapted loss function, the network becomes more sensitive to errors of shadow regions. Thus the proposed network can focus on the learning of robust shadow features. Furthermore, a shadow refinement method is proposed to optimize the boundary region of shadows. In the experiments, the proposed methods are extensively evaluated on two popular datasets and shown better performance on shadow detection compared with current methods.