Spectral-spatial classification of n-dimensional images in real-time based on segmentation and mathematical morphology on GPUs

The objective of this thesis is to develop efficient schemes for spectral-spatial n-dimensional image classification. By efficient schemes, we mean schemes that produce good classification results in terms of accuracy, as well as schemes that can be executed in real-time on low-cost computing infrastructures, such as the Graphics Processing Units (GPUs) shipped in personal computers. The n-dimensional images include images with two and three dimensions, such as images coming from the medical domain, and also images ranging from ten to hundreds of dimensions, such as the multi- and hyperspectral images acquired in remote sensing. In image analysis, classification is a regularly used method for information retrieval in areas such as medical diagnosis, surveillance, manufacturing and remote sensing, among others. In addition, as the hyperspectral images have been widely available in recent years owing to the reduction in the size and cost of the sensors, the number of applications at lab scale, such as food quality control, art forgery detection, disease diagnosis and forensics has also increased. Although there are many spectral-spatial classification schemes, most are computationally inefficient in terms of execution time. In addition, the need for efficient computation on low-cost computing infrastructures is increasing in line with the incorporation of technology into everyday applications. In this thesis we have proposed two spectral-spatial classification schemes: one based on segmentation and other based on wavelets and mathematical morphology. These schemes were designed with the aim of producing good classification results and they perform better than other schemes found in the literature based on segmentation and mathematical morphology in terms of accuracy. Additionally, it was necessary to develop techniques and strategies for efficient GPU computing, for example, a block–asynchronous strategy, resulting in an efficient implementation on GPU of the aforementioned spectral-spatial classification schemes. The optimal GPU parameters were analyzed and different data partitioning and thread block arrangements were studied to exploit the GPU resources. The results show that the GPU is an adequate computing platform for on-board processing of hyperspectral information.

keywords: graphics processing unit (GPU), GPGPU, wavelet transform, denoising, math- ematical morphology, morphological profile, feature extraction, image classification, image segmentation, remote sensing, hyperspectral imaging