葉梅珍Yeh, Mei-Chen陳奕寧Chen, Yi-Ning2019-09-052018-08-272019-09-052018http://etds.lib.ntnu.edu.tw/cgi-bin/gs32/gsweb.cgi?o=dstdcdr&s=id=%22G060547079S%22.&%22.id.&http://rportal.lib.ntnu.edu.tw:80/handle/20.500.12235/106513本論文旨在幫助社群媒體使用者節省大量的時間在選擇照片的濾鏡。由於濾鏡數量的增加，以及手機板面上的限制，要如何快速地選出適合的濾鏡成為一個問題。我們觀察社群媒體上的照片發現，特定的物件與場景會偏好特定的濾鏡，因此希望藉由照片的內容來推薦適合的濾鏡。在本研究中，我們從社群媒體Instagram上蒐集了大量套過濾鏡的照片作為訓練資料，藉由深度學習的技術，分析照片中出現的物件、所在的場景以及美學相關的屬性，建置出推薦照片濾鏡的類神經網路模型。我們在濾鏡推薦的資料集FACD上達到了Top-1 51.87%的準確度，以及從Instagram建立濾鏡資料集，可以讓後續相關的研究使用。This thesis aims to help web users save time on selecting photo filters. Due to the increasing number of photo filters and the limited display size of a mobile phone, filter selection has become an important problem. We observed from the social media sites that photos with specific objects and scenes would prefer certain filters. Therefore, we propose to recommend filters by analyzing the photo content. We collect 68,400 fil-tered photos from Instagram to be used as training data, and analyze the objects, scenes and aesthetics-related attributes from the photos through deep learning tech-niques. We develop a neural network model to recommend photo filters and build a filter photo data set from Instagram to facilitate future research. Experimental results using FACD show Top-1 51.87% accuracy.濾鏡照片內容卷積神經網路Photo FilterConvolutional Neural NetworkPhoto Content