一個自動化葉片辨識系統

Abstract

在郊外踏青總會接觸到許多植物，而讓遊客利用影像辨識了解它們勢必是個趨勢。葉片是植物重要元素之一，因此常用來當作植物辨識的標的物。在本篇論文中，提出了一個針對葉子區域本身為對象的兩階段葉片自動辨識系統。首先，我們萃取出五個特徵植，依據這五個特徵進行辨識動作。在第一階段的辨識中，會將形狀與輸入影像差異大者刪除。接著，便會依據第一階段過濾後的數種名單進行最後的辨識動作。該系統會依照辨識出來的結果提供參考影像，並允許使用者決定最正確的答案。在前五名的準確率為95.14%。此外，我們亦會在前處理時，將拍攝所造成的反光雜訊以及葉片旋轉問題去除。

When wandering around the field, we can touch many plants. It is useful knowing them through image recognition technique. Since leaf is one of the important features for characterizing various plants, it is often taken for plant recognition. The thesis proposes a hierarchical automatic region-based method for leaf recognition. First, delete impossible species to which the input leaf belongs according to the leaf shape represented by five extracted features. Next, based on these candidates, the system finds out the most similar images in our database and allows each user to choose the correct one. The precision rate is 95.14% for top 5. In addition, the proposed method is rotation invariant and solves the noises caused by light reflection in preprocessing.