A Deep Learning Framework to Remove the Off-Focused Voxels from the 3D Photons Starved Depth Images

Photons Counted Integral Imaging (PCII) reconstructs 3D scenes with both focused and off-focused voxels. The portions do not contain or convey any visually valuable information are therefore redundant. In this work, for the first time, we developed a six-ensembled Deep Neural Network (DNN) to identify remove voxels from conventional computational integral imaging PCII techniques. As preprocessing step, used standard Otsu thresholding technique obvious unwanted background. We then preprocessed data train proposed six ensembled DNNs. results demonstrate that methodology can efficiently discard points reconstruct focused-only scene an accuracy of 98.57%.