Recovery and identification of moments in images

• Autor(a) principal: Silva, Júlio Miguel Braz da Costa
• Data de Publicação: 2020
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/30553
• topic: Computer vision; Natural language processing; ImageCLEF; Lifelogging; Moment retrieval

In our modern society almost anyone is able to capture moments and record events due to the ease accessibility to smartphones. This leads to the question, if we record so much of our life how can we easily retrieve specific moments? The answer to this question would open the door for a big leap in human life quality. The possibilities are endless, from trivial problems like finding a photo of a birthday cake to being capable of analyzing the progress of mental illnesses in patients or even tracking people with infectious diseases. With so much data being created everyday, the answer to this question becomes more complex. There is no stream lined approach to solve the problem of moment localization in a large dataset of images and investigations into this problem have only started a few years ago. ImageCLEF is one competition where researchers participate and try to achieve new and better results in the task of moment retrieval. This complex problem, along with the interest in participating in the ImageCLEF Lifelog Moment Retrieval Task posed a good challenge for the development of this dissertation. The proposed solution consists in developing a system capable of retriving images automatically according to specified moments described in a corpus of text without any sort of user interaction and using only state-of-the-art image and text processing methods. The developed retrieval system achieves this objective by extracting and categorizing relevant information from text while being able to compute a similarity score with the extracted labels from the image processing stage. In this way, the system is capable of telling if images are related to the specified moment in text and therefore able to retrieve the pictures accordingly. In the ImageCLEF Life Moment Retrieval 2020 subtask the proposed automatic retrieval system achieved a score of 0.03 in the F1-measure@10 evaluation methodology. Even though this scores are not competitve when compared to other teams systems scores, the built system presents a good baseline for future work.