Research On Optical Image Retrieval Method Based On Shape Feature

Remote sensing imagery has great importance for reconnaissance, precision attack and civil activities, so it has good application prospect to study feature extraction and target recognition methods of remote sensing imagery. The research subject of this dissertation is remote sensing imagery. This dissertation investigates the feature extraction and target recognition methods by focusing our research work mainly on their shape feature.For array targets and blob targets, this dissertation uses recognition method based on data-driving from bottom to up. According to the distribution of gray-scale for blob targets and array targets, this dissertation researches three methods including method based on fuzzy threshold value segmentation,fuzzy threshold segmentation method based on genetic algorithm and ostu method. Experiments prove that ostu method has advantages such as better segmentation result and short execution time for segmentation of the above two kinds of objects. Because of influence of many factors, Segmented result may have many speckles and blockette regions. So this dissertation gives method of removing those regions that isn't object region based on threshold value of region area.For harbor targets, this dissertation uses recognition method based on knowledge-driving. Each port facilities have their unique structure and layout, so the each inner harbor has a special shape. So this dissertation recognizes harbor targets by using its shape of inner harbor. A method of harbor recognition based on its shape of inner harbor. First sea area was separated from land area; then the coastline can be recognized; Finally the inner harbor can be segmented according to closeness measure of character points. Experiments prove that inner harbor can be segmented perfectly by using this method.The object region can be got after image segmentation. This dissertation uses moment invariants as the characteristic vector of object region. This dissertation researches three moment invariants including Hu moment invariants which has translational invariance,rotation invariance,proportional invariance, linear moment invariants, affine moment invariant. This dissertation computes similarity measure between measured image and target template by using normalized euclidean distance. Experiments prove that higher retrival success probability can be got by using the front four Hu moment invariants and affine moment invariant.