Banca de DEFESA: ÁLVARO AMORIM DE ALBUQUERQUE

Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.
DISCENTE : ÁLVARO AMORIM DE ALBUQUERQUE
DATA : 31/07/2026
HORA: 14:00
LOCAL: online
TÍTULO:

Análise de padrões ordinais em imagens para caracterização de ruídoCaracterização e Discriminação de Ruídos em Imagens Digitais Utilizando Planos de Informação e Padrões Ordinais Bidimensionais

PALAVRAS-CHAVES:

Processamento de Imagens. Caracterização de ruído, Padrões Ordinais, Teoria da Informação

PÁGINAS: 50
RESUMO:

Digital images are frequently corrupted by noise during acquisition and transmission, which alters original data and degrades information quality. Effective image restoration is highly dependent on identifying the specific nature of this noise, yet in real-world scenarios, the noise origin is often unknown, significantly hindering the recovery process. This work presents a systematic framework to characterize and discriminate between four common noise types: Gaussian, Salt-and-Pepper, Speckle, and Poisson.
The methodology utilizes two-dimensional ordinal patterns to extract five information-theoretic metrics: normalized Shannon entropy (H), Jensen-Shannon statistical complexity (C), Fisher Information Measure (FIM), Smoothness (τ), and Curve Structure (κ). These measures are jointly analyzed across three representation spaces: the Complexity-Entropy Causality Plane (CECP), the Fisher-Shannon Information Plane (FSIP), and the Smoothness-Curve Structure Plane (SCSP). To validate the characteristic regions of each plane, synthetic images representing periodic, stochastic, and chaotic dynamical regimes were analyzed across multiple embedding windows and image sizes. The framework was then applied to 57 real images from texture (Sintorn, Brodatz) and medical (DPLDs) datasets under two conditions: "Image+Noise," where metrics are computed directly on the corrupted image, and "Isolated Noise," where Total Variation (TV) decomposition first extracts the noise component for independent analysis. Statistical separability between noise types was assessed using Friedman and Nemenyi tests, while within-noise trajectory reproducibility was evaluated via the discrete Fréchet distance and permutation tests.
Planar validation confirmed that each representation space successfully identifies distinct dynamical regimes, though the CECP and FSIP are sensitive to the choice of embedding window and image size through the ordinal pattern sampling criterion — when this criterion is violated, chaotic and stochastic regimes collapse into indistinguishable clusters. The SCSP proved structurally invariant to these choices, consistently separating periodic orbits confined to triangle edges, canonical chaos occupying a negative-κ interior band, and stochastic signals near the origin, independent of image resolution. In the statistical analysis, near-universal noise-type separability was achieved: Gaussian–Poisson was the most consistently distinguished pair in the "Image+Noise" condition across all datasets and planes, while Speckle–Poisson was the least separable pair under all conditions, reflecting the shared Gaussian-like statistical character of both noise types. The SCSP provided a unique diagnostic channel, perfectly separating Speckle from Salt-and-Pepper in isolated signals across all four datasets. Reproducibility experiments confirmed that all noise types trace trajectories significantly more consistent than random chance (p = 0.0 for all 36 permutation tests), with Speckle emerging as the most reproducible due to its multiplicative structure, and Salt-and-Pepper showing the lowest consistency in the "Image+Noise" condition owing to its pixel-replacement mechanism.
These results demonstrate that the three information planes provide complementary and non-redundant diagnostic channels for noise structure analysis. Together, they establish concrete criteria for embedding window selection based on image size and offer a reproducible ordinal-pattern framework for noise characterization in image processing pipelines.

MEMBROS DA BANCA:
Presidente - 1916534 - FABIANE DA SILVA QUEIROZ
Interno(a) - 1647956 - ANDRE LUIZ LINS DE AQUINO
Externo(a) à Instituição - EDUARDO FREIRE NAKAMURA