| 摘要: | 紅外線影像尋標系統 (IIR) 前瞻技術研究的影像融合技術為三年期的研究計畫。研究主題為：第一年發展可見光影像與紅外線影像的融合技術，比較空間域與頻率域影像融合技術的優劣，並實做以小波為基礎的影像融合法則。第二年發展上述光學影像與合成孔徑雷達 (synthetic-aperture radar, SAR) 影像融合技術，並探索合成孔徑雷達影像的特性。第三年則發展上述融合影像後的目標特徵辨識。影像融合的目的是將多張有相關性但具有不同內涵 (content) 的影像融合在一起，將各種內涵同時呈現在一張影像中，讓下一步驟的判斷與分析能夠有更多資訊的支援而得到更正確的結果。影像融合技術有很多種，最基本的方法是高通濾波法 (high-pass filtering method)；最近熱門的方法有：以小波轉換 (DWT) 為基礎的方法、均勻理性濾波法 (uniform rational filter bank)、及拉普拉斯金字塔結構法 (laplacian pyramid)。影像融合技術的應用方式有：不同時間相同地點相同性質影像的融合、相同地點不同解析度的影像融合、不同感測器的影像融合、部份重疊地點的影像融合、..等。不同影像融合的應用，其研究重點也有所不同。例如，有些應用是要從多頻譜影像中的頻譜資料做地表材質 (material) 的分析；因此這樣應用的研究重點在於不破壞原始影像的頻譜資訊 (spectral information)。另外如醫學影像、不同解析度影像、及不同清晰度影像的融合應用，其主要目的是提供人眼觀看；因此這些應用不在乎頻譜資料的改變，只在乎能否保持各自完整的資訊。上述大部份影像都是被動式成像設備所取的，合成孔徑雷達 (SAR) 是一種主動式成像設備，它所取得的雷達影像具有主動穿透偵搜的能力；因此在第二年的計畫中，我們將探索合成孔徑雷達影像的特性，並且發展被動式光學影像與主動式合成孔徑雷達影像的融合技術。經由上述融合後的影像可以提供給我們更多資訊，以利後續的判斷與分析。而本研究的最終目的即是就上述的融合影像執行目標特徵辨識。目標辨識的意義是以空載飛機上的合成孔徑雷達取得雷達影像，再與中波段的被動式紅外線影像融合；最後就融合後的影像執行地面目標辦識。合成孔徑雷達影像較易取得，但中波段的被動式紅外線影像資料較不易取得，因此計畫中將會以一般空載紅外線影像融合、辨識。這也是我們第三年計畫的內容。過去十多年來，我們一直都在從事衛星影像處理、地形景觀分析、及 3D 地理資訊系統相關的研究。已完成與本計畫相關的技術有：利用馬可夫隨機場做多光譜遙測影像分割、非督導式多頻譜遙測影像紋理分割、以多頻譜遙測影像合成紅外線影像景觀、多頻譜影像融合與紅外線影像合成、以關聯隱藏馬可夫樹模式做多重解析度紋理分割、以多時段的SPOT 衛星影像做雲層自動去除等；因此我們有信心來完成本計畫的執行。 ; This study of target detection in infrared images is a three-year research project. The studying topics of the three years are described as follows, respectively. (i) In the first year, we will develop the image fusion techniques for visible-spectral and infrared images. At least, we will implement a wavelet-based image fusion method. We will also survey and compare the spatial-domain and frequency-based fusion methods to provide the useful comments. (ii) In the second year, we will study and survey the properties of synthetic-aperture radar (SAR) images; then we will develop the image fusion algorithm for infrared and SAR images. (iii) In the third year, we will develop the target detection algorithms for the infrared, SAR, and the fused images. The purpose of image fusion is to combine several related but having different content images into an image such that all the contents can be appeared in one single image to benefit the following processing and analysis. There are many different kinds of image fusion methods. The most famous method is the high-pass filtering method; the most hot methods include the wavelet-based fusion method, the methods based on the uniform rational filter bank), and the methods based on the laplacian pyramid image structure. There are also many different applications for image fusion; such as fusing the different template but the same location and same property images, fusing the different spatial resolution images, fusing the different-sensor images, fusing partial out-of-focused images, etc. Different applications have different key studying topics. For example, one application is to analyze the materials of the land cover from the multi-spectral information, the most important thing for such an application is the fusion method can not distort the components of the original spectral information. The application of the medical images is to provide the visible inspection by doctors; thus, we only concern whether all visible information is preserved in the fused images; we don’t care whether the spectral information is changed. All above images are captured from the passive devices. The synthetic-aperture radar (SAR) images are captured from the active radar; such devices can detect partial non-visible data. Thus in the second-year project, we will survey the property of SAR images and then develop the fusion method for the passive infrared images and the active SAR images. From above fusion procedure, the fused images can provide more information for our following detection and decision. The ultimate goal of this research project is to detect targets from any possible sensor. The meaning of target detection is using the aerial-born SAR sensor to get SAR images and using middle-distance passive infrared sensor to get the infrared images. At last, we combine these two kinds of images for land target detection. The land target detection is our studying goal of the third-year research project. In the past years, we have completed several related works, such as the multi-spectral image segmentation, infrared image synthesis, wavelet-based texture segmentation, cloud removal, terrain visualization, etc. Thus we are confident to complete the execution of the research project. ; 研究期間 9801 ~ 9812 |
