Modified U-Net based 3D reconstruction model to estimate volume from multi-view images of a solid object

Abstract

ABSTRACT This paper proposes a modified U-Net based reconstruction model to automatically estimate the volume of solid objects. Initially, the captured multi-view images are pre-processed using improved adaptive weighted mean filter (IAWMF). Then, the pre-processed image is segmented using modified U-Net based learning mechanism to determine ROI. After segmentation, the key feature points are extracted using Harris corner detection approach. Further, the extracted feature keypoints are matched using Oriented FAST and Rotated BRIEF (ORB) approach and the similarity among two images is calculated using similarity detection algorithm. Then, the multiple image views from various angles are used to calculate the depth map using VGGNet-16 to attain the appropriate volume of the object. Finally, the 3D point clouds are reconstructed from multiple images using dense point cloud generation algorithm for volumetric analysis. Finally, the result shows that the proposed model provides better accuracy with less error than other 3D CAD model.