基於字彙樹的大規模書籍影像盤點系統

A Large-scale Image-based Bookstore Inventory Check System Using Vocabulary Tree

莊明翰(Ming-Han Chuang)

指導教授 ： 劉震昌

國立暨南國際大學/科技學院/資訊工程學系/碩士(2012年)

摘要

圖書館和書店都是具有大量藏書的地點，隨著藏書的規模越大也會使得書籍的流動量變大，因此書店往往需要藉由大規模的盤點才能掌握目前書籍的庫存情況。為了盤點的完整性，盤點期間書店是無法營業的，所以這段時間會花費大量的人力與金錢成本。
本論文提出一個自動化的大規模書籍影像盤點系統，在自動拍攝影像的部分，我們使用 AR-Drone 1.0(室內直升機)拍攝影像，並進行書背影像切割與比對實驗。在大規模影像的實驗中，我們擷取出書籍影像的SIFT (Scale-Invariant Features Transform)特徵，並使用窮舉法、字彙樹，及綜合上述兩種方法的 Hybrid 實驗進行影像比對實驗。為了加快窮舉法的比對速度，我們在 Hadoop 平台上將資料庫影像的 SIFT 特徵藉由階層式 K-means 分群建立字彙樹；使得查詢影像與資料庫影像的特徵可以量化成影像字向量 (Visual Words Vector)來加快影像比對時間。此外，本論文在字彙樹與窮舉法上設立門檻值，使得系統可以藉由門檻值來分辨查詢影像是否為資料庫中的影像。
由於每次階層式 K-means 分群產生的字彙樹都不相同，因此本文做了五次的大規模字彙樹實驗，書籍辨識率接近7成，窮舉法有8成，而Hybrid實驗則有7成5的辨識率。在自動拍攝影像上，AR-Drone 的影像解析度並不夠好，但仍有6成的辨識率與7成的切割成功率。門檻值實驗則讓我們的系統可以分辨查詢影像是否為資料庫影像。

關鍵字

書籍影像盤點 ； SIFT特徵 ； 雲端計算 ； 字彙樹

並列摘要

Libraries and bookstores both have a large number of book collections, when the collection becomes greater, the number of checkouts and returning books during the operating hours will also become greater, therefore, bookstores tend to spend a lot of time and manpower for large-scale inventory checking at regular intervals. During the period of large-scale inventory checking, the bookstore is not operating, it takes a great cost for the bookstore.
This thesis proposes an automatic large-scale image-based bookstore inventory checking system. To capture images automatically, we experimentally use the AR-Drone (a kind of indoor helicopter) to take pictures, and then we implement bookspine segmentation and matching in a small-scale database. In the experiments of large-scale image database, we extract the SIFT (Scale-Invariant Features Transform) features from bookspine images, then use exhaustive search, vocabulary tree, and a hybrid method to conduct image matching experiments. Because exhaustive search is time-consuming, we establish the vocabulary tree by hierarchically K-means cluster those database images’ SIFT features on a Hadoop platform. The SIFT features on the query image and the database images can be quantized into visual word vectors by using the vocabulary tree, and it will reduce the time for image matching. Besides, the thresholds on the matching scores are determined when using vocabulary tree and exhaustive search, which make our system capable to determine whether the query image is in the original database or not.
Because K-means clustering would produce a different result at each run, five large-scale experiments were conducted. Book recognition rate achieved 80 percent for exhaustive search, 70 percent for vocabulary tree and 75 percent for the hybrid method. For automatic image capturing using AR-Drone, recognition rate in the small-scale experiment can achieve 60 percent even the quality and resolution of the captured images are low.

並列關鍵字

bookstore management ； SIFT features ； cloud computing ； vocabulary tree

參考文獻

[1] D. G. Lowe, “Object Recognition from Local Scale-Invariant Features,” Proceedings of the Seventh International Conference on Computer Vision, vol. 2, pp. 1150-1157, Sept. 1999.

[2] E. Taira, S. Uchida, and H. Sakoe, “A Model-Based Book Boundary Detection Technique for Bookshelf Image Analysis,” Kyushu University, Japan.

[3] D. Nistér, and H. Stewénius, “Scalable Recognition with a Vocabulary Tree,” Computer Vision and Pattern Recognition, vol. 2, pp. 2161-2168, 2006.

[4] H. Jegou, M. Douze, and C. Shmid, “Packing Bag-of-Features,” Proceedings of the Twelfth International Conference on Computer Vision, 2009.

[5] B. Y. Chou, “An Image-Based Bookstore Inventory Check System,” Master Thesis, National Chi Nan University, 2009.