Infrared Cloud Image Retrieval Technology Research Based On Multi-feature

Infrared cloud image is used to monitor and forecast weather as a kind of weather datum. It ishelpful for weather forecast if the infrared cloud images which are similar to the current infraredcloud image can be found from historical map storage. We research on the technique ofmulti-feature-based infrared cloud image retrieval. We carry out these four jobs: clouds’segmentation, feature’s extract, multi-feature fusion and design of retrieval system.According to meteorology, clouds of the infrared cloud are segmented into four classes whichhas its own characteristics. So segmenting the clouds into four classes is helpful for promoting theretrieval accurancy. We use fractal dimension algorithm to segment the clouds, since differentclass of cloud has different fractal dimension. But the result of the segmentation can’t be accurat,since the earth’s surface has the similar fractal dimension with some clouds. To solve this problem,we we use Otsu algorithm to separate the earth’s surface to ensure the accuracy of results ofclouds’ segmentation. Based on the clouds’ segmentation, we exact gray feature, texture featureand shape feature of every four class of cloud. We use Block-FCM gray histogram algorithm toexact gray feature,making up flaws of tradional ray histogram algorithm, which are the hard-determined bins’ boundary and lacking of the space information, Since the GLCM is not suitablefor infrared cloud images, we use Gabor wavelet-transform algorithm to exact texture feature. Butthe retriecal efficiency will be reduced since the calculated mount of the two-dimension Gaborfilter. To solve the problem, we resolve the two-dimension Gabor filters into two one-dimensionfilters, so that the retrieval efficiency promotes without reduce the accuracy of retrieval outcomes.We use both outline-based method and area-based method to exact shape feature. We choose Cheninvariant moment algorithm as outline-based method, using seven moment invariants as clouds’outline feature. And we also choose geometric invariant moment algorithm as outline-basedmethod, using seven moment invariants as clouds’ area feature. According to the traditionalgeometric invariant moment algorithm, every pixel is thought to has same effect when calculatingcentral moment. But the closer the pixel to the interior of object, the more effect it has whencalculating central moment. To this problem, we use the method of weighting to improve thetraditional geometric invariant moment algorithm, so that the accuracy of retrieval outcomes isimproved. One of image’s feature can only express some of the image’s propertys. So only usingone of image’s feature to carry out the work of image retrieval cann’t usually get ideal outcome.To solve it, we use Bayesian theory to fuse these three features, so that the accuracy of retrievaloutcomes is improved.Based on the frame of content-based image retrieaval system, we use the algorithms above todesign a system of infrared cloud image retrieval based on multi-feature, including the module ofuser interface, the module of image preprocessing, the module of image feature exact, the moduleof database manipulation and the module of feedback of retrieval outcome.