Instagram濾鏡推薦系統

Abstract

近年來，以Instagram為首的照片分享社群平台越來越受用戶喜愛，也因此越來越多的影像濾鏡隨之增加。然而，以Instagram為例，就有超過30個濾鏡供使用者選擇，導致使用者需要花更多時間一一看過濾淨效果、對於濾鏡的選擇也更為困難。因此，我們希望能透過圖像場景分類的幫助，提供使用者一個濾鏡推薦系統。此系統能夠自動辨識影像中之內容，進而推薦五個最適合此場景的濾鏡給使用者，不僅幫助使用者節省濾鏡選擇的時間，也讓其能擁有更佳的用戶體驗。此外，我們也提供了一個新的方法去實作出濾鏡中的暗角效果。一般的暗角效果常常預設照片的重點就在正中央，但我們結合圖像顯著性檢測取得照片中真正的重要部分，並套用修改過的暗角效果去凸顯此重點，幫助觀看者更能在第一眼就能清楚看到照片重點。

With the growing use of social photo-sharing apps such as Instagram, more and more filters are provided for users to create different effects of their photos. However, taking Instagram as an example, it has more than 30 filters, making it quite difficult for users to go through all of them and make a final decision. Therefore, we aim to provide a filter recommendation system for users with the help of scene labeling. This system can automatically detect and understand the content of an image, and list top 5 recommend filters to the users, not only saving their time, but also make the user experience more pleasant. Besides, we also provide a novel way to improve vignetting effect. Instead of generally assume the focus of an image being at the center, we use saliency detection to find out where the focus really is, and highlight it with our modified vignette, helping users spot that object more easily.