| 摘要: | 科技日新月異，隨著電腦硬體技術不斷提升，影像辨識的技術也不斷精進，對於電腦來說，分辨一張圖片的內容是一隻狗或一隻貓，已經是一件非常簡單的事情。然而，要達到高精準度的辨識，需要的條件也有許多：具有良好運算能力的GPU、多至數千與數萬筆的資料、訓練與調整參數優化的時間。在人工智慧技術不斷普及的現在，各個不同領域皆需要利用機器學習與深度學習的技術，達到各個領域所需求的效果，但在學術以外或是較為冷門的領域需要結合人工智慧時，資料量不足首當其衝的浮現了。除此之外，許多產業擁有的機器並不需求運算效能優異的GPU輔助訓練。
目前具有高精準度的影像辨識方法主流，通常仍是以CNN為主的架構，因此需要良好的GPU運算能力與一定的訓練時間才能夠成功訓練出來，雖然也有傳統特徵提取方法結合了類神經網路的PCANet，但在此部分仍有相當大的進步空間。本論文將採用與PCANet相似的架構，同樣是使用傳統方式設計濾波器的方式，將PCA的部分替換為最近特徵線策略NFL，NFL的特性為：在資料量少時能夠保持非常不錯的精準度，利用與PCANet相似的架構進行圖片的分析與處理，並使用NFL提取出必要的特徵，在最後使用SVM方法進行圖片的分類，以上是本篇論文的核心。
分析實驗結果得知，在資料量較少，約500~1000筆左右資料的資料數訓練時，NFLENet能夠得到比PCANet高5%~10%的辨識精準度，並因為資料量減少，訓練時間也大幅減少。
;With the continuous improvement of computer hardware technology, the technology of image recognition is also constantly improving. For computers, it is a very simple matter to distinguish the content of a picture as a dog or a cat. However, to achieve high-accuracy identification, there are many conditions required: GPUs with good computing power, up to tens of thousands of training data, time of training. Nowadays, with the increasing popularity of artificial intelligence technology, different industries need to use machine learning and deep learning to achieve the desired target. However, when there is a need to combine artificial intelligence in areas other than academic or relatively unpopular, the amount of data is insufficient. In addition, many industry-owned machines do not have GPU-assisted training with superior computing performance.
At present, the mainstream image recognition method with high precision is still CNN-based architecture. It requires good GPU computing power and a certain training time to be successfully trained. Although there are also traditional feature extraction methods combined with PCANet based on neural networks. However, there is still big space for improvement in this section. This paper will use a similar architecture to PCANet, but replace the PCA part with the nearest feature line embedding(NFL). The NFL features a very good accuracy when the amount of data is small, and uses a similar architecture to PCANet for image analysis. It is the core of this paper to deal with and use the NFL to extract the necessary features and to use the SVM method to classify the images.
According to the analysis results, NFLENet can obtain 5%~10% higher recognition accuracy than PCANet when the amount of data is small, about 500 pieces of data training, and the training time is greatly reduced because of the reduced amount of data. |
