Defesa de Dissertação de Mestrado: Semi-Supervised Learning Based on Consistency and Contrastive Regularizations For Image Semantic Segmentation Tasks of an Environmental Protection Area

Orientadores:
Gilson Antônio Giraldi - Laboratório Nacional de Computação Científica - LNCC
Mauren Louise Sguario Coelho de Andrade

Banca Examinadora:
Gilson Antônio Giraldi - Laboratório Nacional de Computação Científica - LNCC (presidente)
Roberto Pinto Souto - Laboratório Nacional de Computação Científica - LNCC
Jefersson Alex dos Santos - University of Sheffield

Suplentes:
Fabio Andre Machado Porto - Laboratório Nacional de Computação Científica - LNCC
Rodrigo Minetto - Universidade Tecnológica Federal do Paraná - UTFPR

Resumo:

Semantic image segmentation is a fundamental task in computer vision, aiming to assign a semantic category to each pixel in an image so that spatial information can be extracted in an organized manner. However, supervised semantic segmentation presents a critical limitation: it requires large datasets with pixel-level labels for both training and model validation. This necessity entails a significant investment of time and the work of specialized experts for manual annotation, making the entire process expensive, slow, and challenging to scale. To address this limitation, recent research has turned to semi- supervised learning, which seeks to incorporate unlabeled data into the training process and thereby reduce the reliance on exhaustive manual labeling. Among these approaches, consistency-based methods have stood out for imposing stability on predictions in the face of input disturbances, while contrastive methods are notable for encouraging the model to learn discriminative representations by contrasting similar and dissimilar samples. Building on these ideas, this work conducts a comprehensive evaluation of several state-of-the-art semi-supervised frameworks that employ consistency-based or contrastive regularization strategies during model training. In addition to applying methods from the literature, we also propose our own framework, inspired by existing models, which introduces a new and more robust strategy for selecting negative pairs, a step that plays a central role in effective contrastive learning. All frameworks are applied to the semantic segmentation of satellite imagery from the APA-Petrópolis (RJ) Environmental Protection Area, emphasizing the importance of developing methods that can support studies in environmentally protected regions where manual annotation tends to be costly. The results highlight the effectiveness of these techniques for our target task and demonstrate their practical viability in scenarios with limited availability of annotated data.