Segmentation-Aided Upsampling for High-Resolution Reconstruction of 3-D MMW Images

Three-dimensional millimeter-wave (MMW) low-resolution (LR) images commonly exhibit issues including noise, nonuniform sampling, and blocky sparsity. These characteristics lead to significant source domain inadaptation, making the straightforward adaptation of upsampling networks for 3-D MMW high-resolution (HR) reconstruction less effectiveness. To circumvent this issue, a preprocessing method based on image segmentation is proposed designed for MMW images, which consists of two stages: image expansion and image segmentation. Initially, a conventional clustering algorithm separates the primary foreground image from the raw data. Following this, the first stage involves a 3-D expansion of the foreground image to preserve the structural integrity of the object under the test to the highest degree feasible, thereby generating an initial LR image. Nevertheless, these LR images do not yet meet the prerequisites for the upsampling network. Therefore, a uniform sparse resampling is conducted in the subsequent image-segmentation stage. In this second stage, standard MMW LR images are constructed from their standard MMW HR dataset using the farthest point sampling technique. The similarity between the initial LR images and these standard LR images is computed to create a ground-truth foreground image. PointNet++ is utilized to learn the structural characteristics of the human foreground and further refines the preprocessed LR images through resampling and background separation. These segmented LR images are then fed into the upsampling network, culminating in the reconstruction of 3-D MMW HR images. In the experimental comparison, PUGeo-Net serves as the baseline model. Compared against cluster-based MSGD K-Means and FCM, the proposed method shows improvements of 63%, 59%, and 60% in reconstruction metrics of Chamfer distance (CD), Hausdorff distance (HD), and Jensen-Shannon divergence (JSD), respectively, thus confirming its efficacy.