Research on the Effective Use of apriori Information in Image Interpretation

为了对卫星和航空图像进行准确的土地利用和土地覆盖分析，以及基于医学图像的自动诊断和手术计划，我们的目标是利用所考虑的主题固有的先验知识建立图像解释框架。在高分辨率图像和高光谱图像的研究中，过于自由的模型有时会误导我们得到不一致的结果。通过引入具有足够有限自由度的模型，可以获得足够的信息，结果如下：受损面积估计：将灾前拍摄的卫星图像作为先验信息处理，采用基于“巧合增强”原理的非线性地图算法对灾后拍摄的图像进行对比，对受损房屋进行变化检测。将该系统应用于2003年12月26日巴姆地震的卫星图像，提取出震害区域。将处理结果提供给日本实地考察小组，并通过实地考察验证了检测结果的有效性。作物含糖量的估算：作物的风味取决于其所含糖、氨基酸、蛋白质等成分的比例。在这项研究中，我们试图从收获前的作物高光谱图像中估计这些数量。在估计系统中，我们将神经网络与粒子群算法（PSO）相结合，对高光谱图像中大量波长的数学运算进行了优化。在此估计中，采集后得到的测量值集作为指导数据。人体皮肤提取指数的推导：我们对短波红外区域的高光谱图像进行分析，利用先验信息有效地检测灾后水中或躺在砾石或地面上的人体受害者。因此，提出了NHI（归一化人体指数），利用人体皮肤固有的光谱信息使皮肤有形。周期性分布植被图像的独立分量分析：作为发展日本农田高光谱图像高空观测的基础研究，我们从实验和理论上考虑利用先验知识有效地利用独立分量分析来分辨超出物理分辨率限制的观测图像中的混合像元。少

For the accurate land-use and land-cover analysis of satellite and aerial images, as well as for automatic diagnosis and surgical planning based on medical images, we are aiming at establishing image-interpretation frameworks using a priori knowledge inherent to the subject under consideration. In the study of high-resolution images as well as hyperspectral images, Models with excessive freedom sometimes mislead us to inconsistent results. Sufficient information can be obtained by introducing models with adequately limited degree of freedom, as shown in the following results :Damaged area estimation : The satellite image photographed before disaster was handled as a priori information, and the image photographed after the disaster was compared in nonlinear map algorithm based on the principle of "coincidence enhancement", and change detection of damaged houses was carried out. This system was applied to satellite images of the Bam earthquake of December 26, 2003, and the damaged area w … More as extracted. The processing result was provided to a Japanese field investigation group, and the validity of the detection result was verified via the field study.Sugar content estimation of crops : The flavor of crops is dependent on the fraction of contents such as sugar, amino acid and protein etc. In this study, we tried to estimate these quantities from the hyperspectral image of the crops before the harvest. In the estimation system, we optimized the mathematical operation among large number of wavelength of the hyperspectral images by combining the neural network with the PSO(particle swarm optimization) algorithm. In this estimation, the set of measured values obtained after the harvest was used as a tutorial data.The derivation of human skin extraction index : We analyzed hyperspectral image of short wavelength infrared region, to make use of a priori information effectively in detecting the human victims in the water or lying on the gravel or ground after the disaster. As a result, NHI(normalized human index) was proposed to make the skin tangible by utilizing the spectral information inherent to the human skin.Independent Component Analysis of Periodically Distributed Vegetation Images : As a basic study for the development of high-altitude observation of hyperspectral images of Japanese agricultural fields, we experimentally and theoretically consider the use of a priori knowledge in the effective use of independent component analysis for resolving mixture pixels found in images observed beyond the limitation of physical resolution. Less