基於局部線性嵌入和支持向量機的藻類識別

Abstract

本研究目的在於利用局部線性內嵌(locally linear embedding, LLE)和支持向量機(support vector machines, SVMs)建立有效的藻類識別系統。 總體上，之前的研究顯示，藻類識別系統的準確率大約是90%，但是對於不規則形狀藻類，辨識的準確率會較低。本研究希望延續楊格年代的研究，有效提高系統對不規則藻類的識別率。本系統軟體部分由Matlab 語言開發，用局部線性內嵌提取特徵值，並利用支持向量機進行自動識別。雖然本系統對於雜質部分的識別率較低，但是在不考慮雜質部分的情況下，系統對於非團聚（包括單胞藻、某種未能確認藻種的非團聚的藍綠藻、扭曲單殼縫藻、盤星藻、星鼓藻、柱胞藻）藻類的識別效果令人滿意，識別率都在80%以上。雖然對於未能辨別藻種的團聚的藍綠藻、平裂藻和微囊藻識別率較低，相比於楊格所開發的系統，識別率均有提高。其次本方法分類精度值κ係數為0.8099，也說明它是一種精度較高的方法。再次，利用計算統一設備架構(compute unified device architecture, CUDA)開發的局部線性內嵌的運行速度可提高約一倍。因而，利用基於計算統一設備架構的局部線性內嵌和支持向量機開發的藻類識別系統確實能有效識別藻類，並且比傳統方法更能節約時間。 本系統依然可以進一步改進。首先，在識別藻類之前去除雜質可有效提高識別精度,這極有可能是因為雜質所對應的點並不構成流形。其次，改進局部線性內嵌也有可能提高識別率。第三，在獲取藻類影像的時候應調整合適的景深和視野以提高影像清晰度。第四，稀釋水樣以避免藻類影像重疊。第五，樣本數量過大時，利用局部線性嵌入計算特徵值是非常耗時的，這時利用計算統一設備架構進一步加速局部線性內嵌的計算速度就是必不可少的。

This article is aimed to construct an effective system to implement algae recognition by using CUDA(compute unified device architecture)-based locally linear embedding(LLE) and support vector machine (SVMs approaches). In general, the previous pattern accuracy of algae recognition system is about 90% but it was lower for the recognition of some algae with irregular shapes in the natural water samples. Continuing Yang’s year study, I wanted to achieve a higher accuracy for the identification of the irregularly shaped algae. We used the images of algae captured from charge-coupled device (CCD) and only considered the algorithmic scheme. The algorithm of algal recognition system was constructed on Matlab. Features of algae were extracted by LLE, a manifold learning method, and then algae was classified by SVM, a classifier. Although the recognition accuracy for the unidentified objects is low, the accuracy for all the other algae is satisfactory. By deleting the unidentified objects first, the recognition rates for Chlorella, unidentified separatedCyanobacteria,Monoraphidium, Pediastrum,Cylindrospermum, Staurastrum are more than 80%. The recognition rates for unidentified agglomerated Cyanobacteria, Merismopedia, Microcystis are obviously lower, but they are still higher than the rates in Yang’s research. Besides, the k coefficient of the accuracy of recognition is 0.8099, which means that our recognition system is a method with high accuracy. Thirdly, LLE based on CUDA does accelerate the calculation. According to the results, this algal recognition system rlied on CUDA-based LLE and SVMs is proved to be more efficient and less time-consuming than the traditional method. Also, LLE with SVMs is better to recognize irregularly-shaped algae than explicit feature extraction method with ANN in natural water body. This system can be improved. First, removal of unidentified objects before classification of algae helps to achieve a higher accuracy rate, probably because the corresponding points of these objects do not lie in a manifold. We may also improve accuracy by modifying the existing LLE. In addition we might be able to adjust the depth of field and visual field of microscope and CCD to obtain clear enough images of appointed algae. Also, we need to dilute the samples to avoid the overlapping of several algae. Besides, it is time-consuming to compute the features if the size of sample set of test set is very large. Hence using CUDA to accelerate the process is essential and effective.