利用網路自動擴增訓練影像資料庫以提升行動裝置上花卉辨識之效能

Abstract

在近年來，花卉辨識逐漸受到大家的重視，幫助人們了解花卉的生 物知識。隨著行動裝置的盛行，我們開發出一個即時的花卉辨識系統。 傳統的花卉辨識系統，直接應用在行動裝置上可能會受限於在速度有 限的無線網路中，傳輸資料的速度很慢，以及需要夠多的訓練資料才 能得到比較好的花卉辨識系統。我們利用緊湊的特徵表示符降低傳輸 量達到即時花卉辨識，且將網路上的照片自動加入到系統中，學習出 更好的花卉分類器並取得更多訓練資料。為了實驗我們系統的效能， 我們使用了一個常被使用的花卉資料庫，以及從 Flickr 上下載約一萬 張的影像，來測試系統的準確度與速度。

Flower (plant) recognition has gained much attention recently for helping users to automatically identify flowers (plants). With the convenience of mobile devices, we develop a real-time mobile-based flower recognition system. Applying the traditional flower recognition frameworks to the mobile environment suffer from (1) the bottleneck of the transmission of the query image from the mobile devices to the recognition servers in the limited wireless network and (2) insufficient training images to build the strong classifiers. We purpose a mobile-based detection framework to reduce the transmission time and the expanding training data framework which crawls the training images from web automatically. We evaluate our mobile-based recognition framework on standard dataset, showing performance competitive with existing methods and fast recognition time in the mobile environment. To evaluate the effective of expanding training images from web, we collect 10k flower images from Flickr and show the significant improvement after expanding.