FSCS with Error Sharing based on Alpha Weights for Improving Accuracy of Reconstructed Images

Abstract

Automatic layer decomposition, primarily in the field of image editing, has garnered substantial interest. The prevalent technique is soft color segmentation. Fast Soft Color Segmentation (FSCS), a novel neural network-based method, has been proposed to accelerate the processing time by learning the iterative optimization process responsible for the slow processing time of traditional methods. However, the reconstructed image–obtained by reconstructing the decomposed layers–does not match the original image in terms of saturation and coloring. Therefore, we introduced post-processing involving error sharing based on alpha weights to FSCS (FSCS-ESAW) to improve the agreement between reconstructed and original images. We define the “alpha weight” as the ratio of each alpha layer value corresponding to each color layer to the total value of each alpha layer. FSCS-ESAW shares the reconstruction error–the error that occurs between the reconstructed image and the original image–with each color layer based on alpha weights, thereby improving the accuracy of each decomposed layer. FSCS-ESAW is characterized by its complete independence from FSCS itself and enables getting more accurate images by adding a simple and lowcost post-processing step to FSCS. Experimental results validated the efficacy of FSCS-ESAW, demonstrating superior agreement between the original and reconstructed images compared to FSCS.