Fast and Efficient Shadow Detection Algorithm via Shadow Decoupling and Reparameterization

Since the number of pixels along shadow boundaries is often significantly smaller than that of pixles within shadow regions, accurately detecting shadow boundary areas is more challenging than detecting pixels within the shadow interior. To improve detection accuracy at shadow boundaries, a novel and efficient lightweight boundary-aware shadow detection algorithm called RBNet was proposed in this paper. During the supervised training phase, the input image was divided into shadow and non-shadow regions, and the distance transforms were applied to decouple the shadow and non-shadow regions into a boundary part and a body part respectively. Secondly, the boundary features of shadow regions were learned and the impact of shadow boundaries and interior pixels were balanced in the loss function of RBNet. Additionally, a multi-branch fusion structured re-parameterization module named RepConv was designed in RBNet. Through re-parameterization, the model parameters and computational cost were reduced, and the inference speed was improved. A series of shadow detection comparison experiments and algorithm model comparison experiments between the proposed RBNet and other common shadow detection algorithms is conducted in the paper. Experimental results demonstrate that the proposed shadow detection algorithm, RBNet, not only has the smallest model size but also achieves the fastest inference speed, while outperforming existing shadow detection algorithms in terms of BER performance. RBNet is highly applicable to mobile devices. When combined with shadow removal algorithms, it can significantly enhance the accuracy of object detection or segmentation tasks.