Research on edge‐guided disambiguation module for shadow detection

In shadow detection area, deep learning-based methods have achieved outstanding performance. However, the potential ambiguity in images is still main reason for poor performance of algorithms. By analysing relationships among edges, false-positive information, and areas, an edge-guided disambiguation module better proposed. The consists two parts: 1) edge feature used to guide comprehensive area generation different resolution. 2) features applied reduce features. this way, missed detections added first step false reduced second step. More accurate results obtained. experiments are proceeded on three public datasets: SBU, UCF, ISTD. experimental demonstrated effectiveness robustness proposed method.