人工智慧與深度學習領域被廣泛應用在各個領域中，舉凡車輛辨識、AlphaGo、AOI自動光學檢測、自然語言處理、圖形識別等。近日閒暇時光，欣賞畫作培養興趣的人日漸增加，以及畫作電子化建檔已成為傳統藝術館的新興趨勢。所以我們提供一個應用程式，藉由擷取畫作的特徵，再透過特徵的相似性進行排序，從而使使用者能獲得資料庫中與其畫作風格相近的作品。 本論文使用卷積神經網路-VGG16架構，其模組的訓練資料為ImageNet大約120萬張圖作為訓練。而本論文將畫作當作測試資料，並取得第一層全連接層的權重當作是每張圖獨立的特徵向量再訓練。並且為了加速系統運作，使用主成分分析(PCA)進行一致性的降維。由於每個人對於畫作的風格具有主觀的意識，因此自行設計出一個評價應用程式(Feedback application)，經由不斷的訓練畫作的權重，讓系統濾掉使用者不喜歡的畫作並且依個人喜好推播出相關畫作。 ;Recent years, artificial intelligence and deep learning are widely used in various fields. For example, vehicle identification, AlphaGo, AOI automatic optical detection, natural language processing, graphic recognition, etc. The number of people visiting the art exhibition at their free time has been increased in recent year,and the digit of paintings has become an emerging trend in traditional art galleries. So we provide an application that captures the features of the paintings and then sorts them by the similarity of the features so that users can obtain works in the database that are similar in style to their paintings.
The propose system uses the convolutional neural network-VGG16 architecture. The module training data is about 1.2 million images of ImageNet as training. In this paper, the paintings are used as test data, and the weights of the first layer of the full-connected layer is obtained as an independent feature vector retraining for each image. And in order to speed up the system operation, principal component analysis (PCA) is used for consistent dimensionality reduction.
Since everyone has a subjective awareness of the style of the paintings, our system designed a feedback application. Through the weights of the continuous training paintings, the system filters out the paintings that the user does not like and pushes the related paintings according to personal preference.
