Improving Depth Image Quality via Super-Resolution and Artifacts Removal: CNN-Based Approach vs. Traditional Methods

Abstract

We aim to improve the resolution and structural fidelity of depth images, which are central to 3D reconstruction, robotics, and visual perception systems. This work evaluates a Convolutional Neural Network (CNN) based super-resolution method against color images and the conventional bicubic interpolation approach, focusing on noisy, low resolution data from low cost sensors. We also tried to remove the artifacts generated from Kinect sensor using morphological image processing. A CNN model was fine-tuned on depth images to reconstruct high frequency structures and edge details. Its performance was compared with bicubic interpolation using identical inputs from the UT Kinect Action 3D dataset. Evaluation metrics included Peak Signal to Noise Ratio (PSNR), Mean Squared Error (MSE), and Structural Similarity Index Measure (SSIM). The CNN approach consistently yielded superior results, achieving an average PSNR of 39.82 dB and MSE of 6.43 outperforming bicubic interpolation (35.20 dB PSNR and 19.54 MSE). Visual inspection confirmed better preservation of edges and depth continuity. Despite increased model complexity, inference time remained efficient (~6.2 seconds on GPU) compared to ~30.7 seconds for bicubic on CPU. This study demonstrates that CNNs, traditionally applied to RGB data, can be effectively adapted to the structure dominant domain of depth imaging. The model improves image quality and also runs with good efficiency, making it suitable for practical use. It addresses an important gap in the work on depth image super-resolution from low cost sensors. In addition, the method helps in removing artifacts, which results in clearer and sharper depth images