Year: 2023
Type: dissertation
Abstract: Ground penetrating radar is a tool for extracting information from the underground area.The data is typically measured along lines when scanning areas.To get more accurate information, one solution is to measure along more lines, but GPR surveys take long time.More images can also be generated by using different interpolation methods on the measured data.Previously, a GAN-based approach was proposed for this.It was shown to achieve higher interpolation accuracy than linear interpolation, but there was still room for improvement.In this study, several methods for interpolating ground penetrating radar images models are proposed.Specifically, interpolation is done with deep latent variable models.These approximate the true data distribution better than GANs and are arguably more stable to train.In this research, interpolation experiments were done on VAE-and diffusion based approaches.VAEs learn latent space that has useful properties for image generation, and interpolating between latent vectors produces meaningful images.In addition to VAEs, diffusion models have also been shown to be able to generate high-quality images, especially in the unconditional image generation.Latent space interpolation is also possible in some of the diffusion models.Therefore, VAE-and diffusion-based models should be usable in generating interpolated images.In the first VAE method, the VAE was trained to generate GPR images with one input and one output image.The interpolation was done in the latent space of the trained model by applying linear or spherical linear interpolation on the latents.In the second VAE method, the left image, the right image, and the position information were combined into a three-image input.Interpolation method using a score-based generative model (also known as a diffusion model) was also studied.This model was trained to generate GPR images, and the interpolation was done in the latent space of the trained model.The interpolation accuracy with using VAE and latent space interpolation was similar to using linear interpolation in the image space.However, linear interpolation had worse accuracy near the edges of the underground objects.With the three-image input approach, the interpolation accuracy was similar to the accuracy of the previous GAN-based method.With diffusion model, the interpolation accuracy was close to the accuracy with VAE using one-image input.However, the accuracy was worse near the edges of objects.So, linear and spherical linear interpolation may be too simple for latent space interpolation.The learned latent spaces of the one-image input and the three-image input methods were also analyzed.In the one-image input method, the latent space was more dispersed compared to the three-image input method.With three images as input, the model was also not suitable for generating more GPR images, unlike the model with one image as the input.This suggests that a model that is trained to both interpolate and generate the interpolated image might generate poor images in some cases.There are various ways to continue this research.Only simulated ground penetrating radar images were used in the experiments.The earlier GAN-based interpolation method was tested on both real and simulated data, and these methods should also be tested on real images.Additionally, the number of images used for training and testing the models was quite small.The methods used in this study should benefit from more data, because the learned latent spaces were not very smooth.With more data, the latent spaces should have less gaps, and especially doing interpolation in the latent space should become more accurate.One possibility is to train a network for calculating the latent space interpolation instead of using linear or spherical linear interpolation.This should provide good image generation performance and interpolation accuracy.However, further research should be done to confirm this.
Author: Klemettinen, Lauri
Language: English
Subfield: Geophysics
Field: Earth and Planetary Sciences
Domain: Physical Sciences
Open Access status: green
Funder: Itä-Suomen Yliopisto
OpenAlex ID: W7215302258
