回饋式影像自動標註與檢索系統

Abstract

以傳統方式建構關鍵字式影像檢索系統時，必須花費大量的人力與時間為影像進行標註，然而標註的內容往往受到標註人員主觀性的影響。本論文提出一套自動影像標註方法，利用影像切割取得影像中的物件，透過共同出現模式，採用下列三項技術協助影像進行標註：1.利用區域式影像切割，將影像切割成幾個跟人們視覺上比較吻合的物件；2.將所取出之物件對映到最接近的前三群進行正規劃取得更適合物件的語意概念；3.加強位於影像中央物件之語意概念所佔的權重。本論文亦透過使用者相關回饋的方法，提供使用者更為精確的檢索結果。由實驗中得知，相較於傳統共同出現模式，本系統平均準確率提升了19.45%，經過五次的相關回饋後，系統的平均準確度可以由原本的46%提升到82.7%。

Conventional keyword-based image retrieval systems take a large amount of human labor to annotate images; furthermore, the annotations given by human are subjective to annotators. This thesis proposes an algorithm to automatically annotate images. This algorithm exploits three techniques to modify the co-occurrence model proposed by [Mori99]. 1. Segment an image into regions, each of which corresponds to an object. The regions identified by region-based segmentation are more consistent with human cognition than those identified by block-based segmentation. 2. According to visual features (color and shape), map an identified region into three most similar clusters to obtain its associated semantic concept. 3. Strengthen the weight of a region located in the center of an image, because the central object is usually the focus of an image. The semantic concepts derived by the above algorithm may not be the same as the real semantic concepts of the underlying images, because the former concepts depend on the low-level visual features. To ameliorate this problem, we propose a relevance-feedback model to learn the long-term and short-term interests of users. The experiments show that the proposed algorithm outperforms the traditional co-occurrence model about 19.5%; furthermore, after five times of relevance feedback, the mean average precision improves from 46% to 82.7%.