Due to factors such as non-uniform human gray scale, occlusion, and camera movement, the results Rt and Rt' obtained by the above LSAC model are likely to have human body fragmentation and noise problems, which need further processing. Image inpainting is a common method to improve the integrity of broken objects. However, they often have difficulties in computational efficiency or parameter selection, and are not practical for human object inpainting. Morphological filtering is also a common method to remove noise, but conventional morphological filters often lead to jagged edges and destroy the smoothness of the target contour output by the LSAC module.
In order to improve the integrity of the human body area and remove noise at the same time, the segmentation results Rt and Rt' are fused based on open reconstruction technology. Open reconstruction starts from the seed point, gradually expands the seed area on the mask image, and finally restores the sub-area where the seed point is located in the mask image, so as to achieve the purpose of filtering out other sub-areas. Compared with conventional morphological filters, open reconstruction can keep the shape of connected regions exactly the same before and after reconstruction. The figure below shows the effect of rebuilding.