| 摘要: | 近年來，隨著硬體的計算能力快速提升，建置成本下降以及大數據的興起，使用機器學習與深度學習技術的應用越來越普及。機器學習與深度學習技術最常運用的用途就是識別與預測。觀察大自然界，一個物體的電磁波頻譜，指的是這物體所發射或吸收的電磁波的特徵頻率分佈。電磁波譜頻率從低到高分別為無線電波、微波、紅外光、可見光、紫外線、X射線和伽瑪射線，在這廣泛的頻譜分布範圍裡，值得我們探討的標的識別與分類的議題很多，其中，我們探索兩個議題作為分別基於機器學習和深度學習方法的識別應用研究。
第一個研究議題是掌紋識別，掌紋通常透過可見光或紅外光進行擷取，相關的研究雖然已經很多，其中，關於非接觸式掌紋識別與多光譜掌紋識別是較新的議題；使用非接觸式設備收集掌紋圖像具有掌紋識別的多個優點，例如: 用戶友善性，衛生性和抗偽造性。而多光譜掌紋由於捕獲的光譜帶不同，多光譜掌紋將獲得不同的特徵。因此，這兩者對研究人員來說更具吸引力。本文提出了一種基於多光譜掌紋識別的多階層融合方法，用於個人識別。首先，不需要關於多光譜圖像的先驗知識，並且可以自動設置所使用的參數。其次，由於使用的是非接觸場景中擷取之掌紋圖像，因此，所提出的方法增加了用戶友善性，安全性和衛生性。第三，它可以恢復非接觸式掌紋圖像的幾何變形，並自動對齊並裁剪掌紋圖像上的感興趣區域。第四，介紹了一種分層融合方案，包括數據級和特徵級的融合。數據級融合使用離散小波變換將感興趣圖像分解為四個波段，並使用反向離散小波變換數據級別融合四個波段的圖像。本文還提出了一種係數合併方案，用於合併由離散小波變換從四個頻帶圖像分解的四個係數矩陣。第五，通過使用加伯濾波器從融合的感興趣圖像中提取基於紋理的特徵。第六，通過使用多分辨率分析獲得多個特徵，該多分辨率分析應用了多個多分辨率濾鏡從融合的感興趣圖像中提取多個特徵。最後，將每張融合的感興趣圖像的高維特徵矩陣轉換為一維特徵向量，該一維特徵向量用作支持向量機分類器的輸入數據。支持向量機用於在特徵級別融合多個特徵，同時它也用作分類器。
第二個研究議題是船艦識別，船艦的特徵信息通常可透過雷達的微波蒐集到相關的回波資訊，目前，可以運用於識別雷達目標的特徵有很多，包括高解析度距離輪廓圖、合成孔徑雷達微波影像、逆合成孔徑雷達微波影像…等。其中，高解析度距離輪廓圖是一種方便又容易運用的目標資訊，它的數據相對較小，因此，基於高解析度距離輪廓圖的雷達自動目標識別技術一直受到相關領域的專家學者關注。現有的研究已存在了許多傳統圖形識別的方法，運用深度學習的方法相對較少，本研究主要的貢獻除了自行蒐集並建構一個真實的高解析度距離輪廓圖船艦資料庫，而且還專注於運用深度學習方法進行船艦目標的識別與分類，包括了卷積神經網路、長短期記憶網路、雙向長短期記憶網路以及本文所提出的使用雙通道卷積神經網路與雙向長短期記憶網路組合的模型。在傳統的雷達高解析度距離輪廓圖目標識別方法中，雷達的先驗知識對於目標識別必不可少。深度學習方法在高解析度距離輪廓圖中的應用始於近年來，並且大多數是卷積神經網絡及其變體，而遞歸神經網路以及遞歸神經網路和卷積神經網路的組合則相對較少使用。雷達發出的連續脈衝擊中了艦船目標，接收到的反射波的高解析度距離輪廓圖似乎提供了艦船目標結構的幾何特徵。當雷達脈衝發送到船上時，船上的不同位置具有不同的結構，因此高解析度距離輪廓圖中反射的回波的每個測距單元將不同，並且相鄰結構也應具有連續的關係特性。這啟發了作者提出了一個模型，以將雙通道卷積神經網路提取的特徵與雙向長短期記憶串聯起來。雙通道卷積神經網路中使用了不同的濾波器來提取更多的深層特徵，並將其饋入後續的雙向長短期記憶網路。雙向長短期記憶網路模型可以有效地保留關鍵信息並實現雙向時序依賴性。因此，相鄰距離單元之間的雙向空間關係可用於模型中作為具鑑別度的識別特徵。實驗結果表明，該方法對船艦識別具有魯棒性和有效性。
;In recent years, with the rapid increase in hardware computing power, the decrease in construction costs and the rise of big data, the application of machine learning and deep learning technology has become more and more popular. The most common use of machine learning technology is recognition and prediction. Observing the natural world, the electromagnetic spectrum of an object refers to the characteristic frequency distribution of electromagnetic waves emitted or absorbed by the object. The electromagnetic spectrum frequencies are listed from low to high as radio waves, microwaves, infrared, visible light, ultraviolet rays, X-rays and gamma rays. Therefore, in this wide spectrum distribution range, there are many topics worthy of research on target recognition and classification. Among them, we explore two topics as recognition application research based on machine learning and deep learning methods respectively.
The first is palmprint recognition. It is usually captured by visible light or infrared light. Although there have been many related studies, contactless palmprint recognition and multispectral palmprint recognition are relatively new topics. Collecting palmprint images using contactless equipment has multiple advantages of palmprint recognition, such as user-friendliness, hygiene and anti-counterfeiting. Multispectral palmprints will acquire different features because of the different spectral bands captured. Therefore, these two are more interesting for researchers. This paper proposes a reliable and robust biometrics based on multispectral palmprint images for personal recognition. Firstly, no prior knowledge about the multispectral images is necessary and the used parameters can be set automatically. Secondly, the palmprint images are captured in contactless scenarios. It is without any docking device. Thus, the proposed approach increases the user-friendliness, security and sanitation. Thirdly, it can restore the geometric deformations of contactless palmprint images, and align and crop the region of interest (ROI) on palmprint images automatically. Fourthly, a hierarchical fusion scheme, including data-level and feature-level fusion, is introduced in this paper. The data-level fusion uses discrete wavelet transform (DWT) to decompose the ROI image into four bands and inverse discrete wavelet transform (IDWT) to fuse four bands images at data level. This paper also derives a coefficient merging scheme to merge the four coefficient matrices which are decomposed by DWT from four band images. Fifthly, texture-based features are extracted from the fused ROI images by using Gabor filter. Sixthly, multiple features are obtained by using multiresolution analysis (MRA) that applies multiple multiresolution filters (MRFs) to extract multiple features from fused ROI images. Finally, the high dimensional feature matrixes of each one fused ROI image are reshaped and concatenated to a one-dimensional feature vector which is used as the input data to support vector machine (SVM) classifier. SVM is used to fuse multiple features at feature level. It also is used as a classifier.
The second is ship recognition. Features information of ships can usually be collected through radar microwaves to collect relevant echo information. At present, we can see that there are many features that can be used to recognize radar targets, such as high-resolution range profiles (HRRP), synthetic aperture radar (SAR) microwave images, and inverse synthetic aperture radar (ISAR) microwave images. Among them, HRRP is very convenient, which has relatively small data. Therefore, radar automatic target recognition (RATR) based on HRRP has always received extensive attention from experts and scholars engaged in RATR research. There are many conventional pattern recognition methods in existing research, and there are relatively few deep learning methods applied in this field.
The main contribution of this research is not only to collect and construct a real-life HRRP ship dataset, but also to focus on the use of deep learning methods to recognize ship targets, including convolutional neural network (CNN), long short-term memory (LSTM), bidirectional long short-term memory (BiLSTM) and the proposed model using a combination of two-channel CNN and BiLSTM. Radar HRRPs describe the radar characteristics of a target, that is, the characteristics of the target that is reflected by the microwave emitted by the radar are implicit in it. In conventional radar HRRP target recognition methods, prior knowledge of the radar is necessary for target recognition. The application of deep learning methods in HRRPs began in recent years, and most of them are CNN and its variants, and recurrent neural network (RNN) and the combination of RNN and CNN are relatively rarely used. The continuous pulses emitted by the radar hit the ship target, and the received HRRPs of the reflected wave seem to provide the geometric characteristics of the ship target structure. When the radar pulses are transmitted to the ship, different positions on the ship have different structures, so each range cell of the echo reflected in the HRRP will be different, and adjacent structures should also have continuous relational characteristics. This inspired the authors to propose a model to concatenate the features extracted by the two-channel CNN with BiLSTM. Various filters are used in two-channel CNN to extract deep features and fed into the following BiLSTM. The BiLSTM model can effectively capture long-distance dependence, because BiLSTM can be trained to retain critical information and achieve two-way timing dependence. Therefore, the two-way spatial relationship between adjacent range cells can be used to obtain excellent recognition performance. The experimental results revealed that the proposed method is robust and effective for ship recognition. |
