基於深度學習之相片美學評分系統;A Deep-learning-based Photo Aesthetics Assessment System

NCUIR > College of Electrical Engineering & Computer Science > Graduate Institute of Computer Science and Information Engineering > Electronic Thesis & Dissertation > Item 987654321/84118

Please use this identifier to cite or link to this item: http://ir.lib.ncu.edu.tw/handle/987654321/84118

Title:	基於深度學習之相片美學評分系統;A Deep-learning-based Photo Aesthetics Assessment System
Authors:	劉軒宏;Liu, Hsuan-Hung
Contributors:	資訊工程學系
Keywords:	深度學習;影像質量評估;視覺藝術;攝影構圖;deep learning;image quality assessment;visual art;photographic composition
Date:	2020-08-18
Issue Date:	2020-09-02 18:06:59 (UTC+8)
Publisher:	國立中央大學
Abstract:	過去，攝影作品的好壞除了與拍攝技術有關之外，攝影機的硬體功能也是其中的一大關鍵。但隨著近幾年智慧型手機的問世與進步，拍攝相片時，其搭配之硬體功能已不再構成作品好壞之關鍵阻礙。然而相較於硬體功能之長足進步，個人的攝影技術卻有因人而異的大差距。除此之外，過去傳統使用底片來攝影時，我們只能將相片洗了之後再評估好壞，現在有許多後製的工具可以使用以讓相片更為吸引人。因此，本論文提出一套能夠評分、改善以及分析構圖之系統，期望在協助使用者改進相片之餘，也能透過分析出的相片構圖，讓使用者理解如此改進的理由，以及如何能夠拍出更好的相片。本論文系統包含（1）利用NIMA模型對輸入之相片給予評分（2）以NIMA模型之給分作為基準，給予相片之裁切調整與顏色等參數調整建議（3）對調整後之相片進行構圖分析。根據系統實驗結果顯示，NIMA之評分結果符合一般人類之普遍審美；構圖分析之平均Top-5正確率達到92.08%。因此，本系統具備一定程度之可用性。;In the past, in addition to the quality of photography, the camera′s hardware function was also a key factor. However, with the advent and advancement of smartphones in recent years, the hardware functions that are used when taking photos are no longer a key obstacle to the quality of the work. However, compared with the great progress of the hardware function, there is a big gap between individual photography techniques. In addition, in the past, when traditionally using negatives for photography, we need to develop a roll of film first before evaluating the quality. Now there are many post-production tools that can be used to make the photos more attractive. Therefore, this paper proposes a system that can assess, enhance and analyze composition. In addition to helping users improve their photos, we hope that the composition analysis of the photos can make users understand the reasons for the improvement and how they can take better photos. This system includes (1) using the NIMA model to assess the input photo (2) giving suggestions for cropping and color adjustment of photos using the score of the NIMA model as a benchmark (3) giving composition analysis of the adjusted photos. According to the results of the experiments, the results of the NIMA model are in line with the general aesthetics, and the average Top-5 accuracy of composition analysis reaches 92.08%. Therefore, this system has a certain degree of usability.
Appears in Collections:	[Graduate Institute of Computer Science and Information Engineering] Electronic Thesis & Dissertation

Files in This Item:

File	Description	Size	Format
index.html		0Kb	HTML	195	View/Open

社群 sharing

Loading...