Accelerated Computational Methods in Image Processing using GPU Computing: Variational Saliency Detection and MRI-CT Synthesis

Abstract

The parallel computing paradigm has made a significant impact on many scientific areas whose algorithms have experienced a substantial improvement. In Computer Vision, there has been an explosion of real-time applications ranging from medical science to telecommunications, defense, aerospace, etc. However, designing and implementing algorithms that make use of this paradigm introduce some challenges. The main goal of this Thesis is to provide efficient algorithms for two relevant computer vision problems capable of real-time resolution based on multicore on CPU and manycore on GPU. The first problem is Saliency detection and the second one is Computed Tomography synthesis. Nowadays, there is a large amount of visual content (videos and images). As humans, our visual attention can discriminate the relevant information of each scene in a fast way that allows our limited resources to interact with the environment naturally. Computer Vision has recently started proposing algorithms to mimic the human visual attention system. In particular, saliency detection belongs to a binary segmentation problem. These algorithms extract the relevant part of the scene automatically. Saliency detection models are of much interest to the research community. These saliency models aim to understand how visual attention works in humans and to be able to discriminate important information automatically and process it adequately. The first contribution of this Thesis is the development of a variational model for the resolution of the saliency detection problem in natural images. Variational methods have a long history in Mathematics and Engineering. Its application to low-level image processing such as optical flow, denoising, inpainting, deblurring, etc. has provided state-of-the-art results for unsupervised methods in the computer vision community. One common drawback of these methods, based on local differential operators, is their inability to handle textures. Non-local variational methods, based on non-local differential operators, have been introduced successfully in image processing to overcome this problem. They model not only proximity but also similarity features. Few algorithms solve the saliency detection problem using the variational setting in the literature. In a non-local framework, given by a 5-dimensional feature space, we propose a new general model based on the non-local Total Variation (NLTV) operator. Different scenarios are explored in the modelling exercise introducing a new explicit term for saliency detection. The resulting models are validated and the related optimization problems are solved on CPU and GPU platforms for comparison. A primal-dual algorithm dictated by the mathematical formulation of the problem is implemented. A comparison with previous variational approaches is presented. The Quantitative results, under typical metrics used in saliency in complex public datasets, indicate that our method obtains almost the best score in all metrics. Furthermore, the implementation of the algorithm either on CPU or GPU achieves up to 33 fps and 62 fps respectively for 300 400 image resolution, making the method eligible for real-time applications such as surveillance video, temporal video classification, background subtraction, object detection, and general unmanned aerial vehicle (UAV) scene image recognition. On the other hand, computer vision and image processing techniques have had a tremendous impact on Medical Imaging during the last few decades. For example, these applications have made it possible to diagnose diseases at an early stage. The research community has focused on multimodal imaging using two or more acquisition techniques to facilitate the diagnostic. It has several advantages because every modality aids the diagnosis process by providing specific information about the anatomy of the body. The combination of data from different modalities can be used for accurate construction of patient-specific tissue models for dosimetry applications in electromagnetics or the use of tissue information for attenuation correction in Positron Emission Tomography and Magnetic Resonance Imaging (PET/MR) among other applications. Methods that allow generating a new modality image from a given data set through image processing are interesting because they reduce acquisition time and sometimes some harmful modality. The second contribution of this Thesis is a novel patch-based approach to generate Computed Tomography volumes (CT) from Magnetic Resonance (MR) data. The proposed method can be useful for several applications such as electromagnetic simulations, cranial morphometry, and attenuation correction in PET/MR. The proposed algorithm implemented using GPU computing techniques improves 15.9 against a multicore CPU solution and up to about 75 against a single core CPU solution. The proposed solution produces high-quality pseudo-CT images when a neighbourhood of 9 9 9 and 10 atlases are used because the patientspecific CT and pseudo-CT images are similar according to the metric normalized cross-correlation (NCC = 0.93). Furthermore, the algorithm has been revisited with new hardware (NVIDIA DGX Station with NVIDIA V100 GPUs) in the last part of this Thesis. The new results have achieved almost 6 speedup in comparison with the best previous configuration. These results have confirmed the scalability of the method in new architectures.