Transformada floor of log aplicada em contexto cross-dimensional

• Ano de defesa: 2023
• Autor(a) principal: Peixoto, Solon Alves
• Orientador(a): Rebouças Filho, Pedro Pedrosa
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Área do conhecimento CNPq: CNPQ::ENGENHARIAS::ENGENHARIA ELETRICA
• Link de acesso: http://repositorio.ufc.br/handle/riufc/76791

This work proposes a grouping method, independent of dimensionality, called Floor of Log (FoL) transform. The advantage of this method is its ease and practicality in implementation, as well as the ability to generate multiple effects on data, among which clustering and compression stand out. Three applications were chosen aimed at solving real problems. In the first, FoL was evaluated in tasks related to focused on facial recognition, specifically on feature arrays. For this assessment, the CelebA, Extended YaleB, AR and LFW datasets were used together with the analysis of the dataset size after applying FoL and accuracy(ACC) of the matching result between the faces. The second application evaluates FoL in a two-dimensional environment, specifically in computed tomography for lung segmentation and, consequently, within an image processing environment. In this evaluation, the LUNA16 and LAPISCO together with Haunsdorff Distance(HD), DICE, ACC, Jaccard and Matthews Correlation Coefficient (MCC) . The third application seeks to evaluate FoL in a context more dimension-independent, within general-purpose convolutional neural networks. The CIFAR10 and CIFAR100 benchmark datasets were used, in addition to Davies-Bouldin(DB), Calinski-Harabasz(CH) and Silhouette (Sil). As a result, FoL when applied to arrays in the CelebA, Extended YaleB, AR and LFW datasets, obtained equal or better results when compared with the approach using the same classifiers with features not compressed, but with an 86 to 91% reduction compared to the original size of the data. In a two-dimensional environment, FoL was applied for lung segmentation in computed tomography images. The FoL algorithm achieves good results with approximately 19 seconds in the most significant result in an exam with 430 slices and presents similarity indices reaching HD 3.5, DICE 83.63, and Jaccard 99.73 and indices qualitative reaching Sensitivity 83.87, MCC 83.08, and ACC 99.62. Finally, the FoL was also presented as a supervised clustering transform that can be trained to achieve better results and attached to other approaches such as Deep Convolutional Neural Networks reaching DB 1.74, CH 137 and Sil 0.17.