Do Macro ao Micro: Abordagens em Reconhecimento e Segmentação de Objetos em Imagens

• Ano de defesa: 2025
• Autor(a) principal: Anderson Aparecido dos Santos
• Orientador(a): Wesley Nunes Goncalves
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Fundação Universidade Federal de Mato Grosso do Sul
• País: Brasil
• Palavras-chave em Português: Drones, Sensoriamento Remoto, Aprendizagem Profunda
• Link de acesso: https://repositorio.ufms.br/handle/123456789/12481

Detecting small objects in high-resolution images is a significant challenge in computer vision, especially in agricultural scenarios, where the size of insects tends to occupy only a few pixels relative to the entire image. Despite the advances brought by Convolutional Neural Networks (CNNs), the standard use of these architectures presents limitations, as the information associated with small objects tends to be lost due to downsampling and pooling operations. In this context, this thesis proposes strategies to improve the detection and segmentation of small objects. Initially, images are divided into smaller patches, and each patch is used individually to train and validate the models. For inference and prediction on the test set, the patches are overlapped, ensuring that in at least one of them, the object remains uncut. To address the issue of objects that end up being divided among different patches, techniques are presented to filter out-of-pattern predictions, those with low confidence scores, or with redundant and overlapping regions. To allow these techniques to also be applied to segmentation approaches, masks were generated from the original annotations, enabling the evaluation of both detection and segmentation models. Considering the limitations of the traditional Intersection over Union (IoU) metric for small objects, especially due to its sensitivity to minor spatial inaccuracies, this thesis also proposes an alternative metric based on the distance between the centers of the bounding boxes. The experimental results demonstrate that the proposed approaches contribute to the localization of small objects in high-resolution images, showing that both detection and segmentation techniques can be effective, as long as the data are properly processed before and after being input into the model.