Managing feature extraction, mining and retrieval of complex data: applications in emergency situations and medicine

• Autor(a) principal: Chino, Daniel Yoshinobu Takada
• Data de Publicação: 2019
• Tipo de documento: Tese
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: https://www.teses.usp.br/teses/disponiveis/55/55134/tde-29012020-142713/
• topic: Bag-of-visual-words; Bag-of-visual-words; CBIR; CBIR; Content-based image retrieval; Detecção de fogo; Fire detection; Image segmentation; Recuperação de imagens baseada em conteúdo; Segmentação de imagens; Skin ulcer; Úlceras cutâneas crônicas

The size and complexity of the data generated by social media and medical images has increased in a fast pace. Unlike traditional data, images cannot be dealt with in its original domain, leading to rising challenges in knowledge discovery tasks. The image analysis can aid on several decision making tasks. Crowdsourcing images such as social media images can be used to increase the speed of authorities to take action in emergency situations. Images taken from the medical domain can support on daily activities of physicians to diagnose their patients. Content-Based Image Retrieval (CBIR) systems are built to retrieve similar images, being an important step for the knowledge discovery. However, in some image domains, only parts of the image are relevant to the problem. This PhD research is based on the following hypothesis: the integration of image segmentation methods with local feature CBIR system improves the precision of the retrieved images. We evaluate our proposals in two images domain: fire detection on urban emergency situations and chronic skin ulcer images. The main contributions of this PhD research can be divided in four parts. First, we propose BoWFire to detect and segment fire in emergency situations. We explore the combination of color and texture features through superpixels to detect fire in still images. Then, we explore the use of superpixels to extract local features with BoSS. BoSS is a Bag-of-Visual-Words (BoVW) approach based on visual signatures. To integrate segmentation methods with CBIR, we propose ICARUS, a skin ulcer image retrieval framework. ICARUS integrate segmentations methods based on superpixels with BoVW. We also propose ASURA, a deep learning segmentation method for skin ulcer lesions. Besides segmenting skin ulcer lesions, ASURA is able to estimate the area of the lesion in real-world units by analyzing real-world objects present in the images. Our experiments show that our proposals achieved a better precision while retrieving the most similar images in comparison with the existing approaches.