隨著商標資料庫的增加，容易遇到商標相似度過高，或是仿冒的問題。但是龐大的商標資料庫如果只依賴人工來進行分類以及相似商標之搜尋、比對的話，勢必會耗費大量的人力資源及時間成本。相似商標之準確度也可能因為作業流程而下降。 本篇論文提出一種混合色彩空間特徵，以及區域紋理特徵兩種方法的商標辨識系統。首先將圖片進行前處理，去除不必要的白色邊緣部份，只留下系統所需要的商標部份，並且對它進行放大縮小、平移這兩種正規化。接著將圖片等分成數個大小相同的區域，計算各個區域中之色彩平均值，此為第一項特徵。第二項特徵是將商標圖片進行二值化的運算，再利用區域二維特徵來表示。 在實驗中，使用者會輸入一張查詢影像，由上述兩項特徵來對所有資料庫中的商標圖案進行相似度的比對。而因為相似度的判定是形狀相似大於色彩相似，所以本方法會給予區域二維特徵較高的權重。 經過多種方法實驗比較之下，本篇論文提出之商標辨識系統在準確度上的表現較其他單一種檢索方法更為優異，因此結果顯示本論文實驗方法具可行性。 With the increasing amount of trademarks in trademark database, one will suffer the problem in retrieving similar or counterfeit trademarks. It is very time-consuming to classify, verify and retrieve the huge amount of trademarks by human intervention. Moreover, the accuracy of classification will decline with time due to human factor. In this thesis, we propose a trademark retrieval system by hybridizing color-spatial features and the local texture features. Firstly, preprocessing of trademark image is performed to remove unnecessary part of the white edge and retain the region of interest that we need. Normalization with scaling and shifting is also accomplished in this step. Then, divide the image into a number of regions and calculate the color mean value of each region, which is the first feature extracted of the retrieval system. The second feature is to binarize the trademark image and use two-dimensional features as local features to represent the image. In retrieval, the user will input a query image and execute the similarity measurement by utilizing the proposed features with all the stored trademarks in the database. Since the discrimination power of shape is higher then color by experimentation, the two-dimensional characteristic region be assigned higher weight in similarity measurement. By conducting various experiments, the accuracy of the proposed trademark retrieval system outperforms other retrieve methods. It demonstrates the feasibility and efficiency of the proposed system in trademark retrieval.
