Shadow detection based on vision sensors is widely used in image processing. Because of the variability of illumination and projection surface color, shadow detection based on a color image is a challenging problem. Aiming at solving the conflict between the complexity and robustness of current shadow detection algorithms, we established a new shadow detection network by combining the global thresholding method with a neural network, which realized the decoupling of the global threshold and binary fusion. Three public shadow detection datasets, large-scale shadow dataset of Stony Brook University (SBU), large-scale dataset with image shadow triplets (ISTD), and shadow detection for mobile robots features evaluation and datasets (SDMR), were utilized for its verification. Experimental results show that the performance of the proposed network approaches that of previous deep learning methods, both visually and in terms of objective indicators, but the proposed network has the advantages of a simple structure and good robustness.