| Object based high resolution remote sensing image object recognition and classification has been the research focus in remote sensing, photogrammetry and GIS. Object-oriented object recognition method need to use the spectral, geometry, texture and spatial neighborhood relation information. Objects are usually got from image through image segmentation. We could find that the segmentation results will directly influence the recognition and classification accuracy. That made the object-oriented high resolution remote sensing image segmentation has been the main research direction of image segmentation. Graph optimal theory based object-oriented image segmentation are presented in this paper. Mainly focus on the minimum spanning tree(MST), considering the character of the remote sensing images and the segment of MST, edge weight, segment criterion and a MST based multi-scale, multi-level image segmentation system are designed. The detailed research work and suggestions can be sum up as the following:1. Different dissimilarity measure based edge weights are designed for MST segmentation. Focuses on the main data source (high resolution multiband images), one edge weight constructing method is proposed based on Euclidean distance considering the band weight and NDVI, which can improve the segment precise of vegetation, buildings and roads. The other one constructs the edge weight with spectral angle mapping (SAM), which has a good segment result in roads, buildings, bare ground and water.2. Two optimal criterion of MST based on the edge weight is expounded. From the point view of fast segmentation and after analysis of MST segmentation methods, proposed two optimal criterions based on the weight property. One is a simple and useful criterion which can be used in any edge weight function. This criterion according to the weight is the reflection of region edge, two pixels with the edge weight less than the threshold should belong to one region and they will fall in the different regions with the edge weight more than the threshold. The less edge is shown in the large scale and more edge retains in the small scale. Through setting the percentage of the edge to be retained and adopt the lag criterion to determine the merge threshold. Experiment shows this criterion is easy to set parameter and keeps more details of the region. The other one is based on the statistics theory, the image segmenting process is considered as a study-predict process. A loss function is constructed with the edge weight. Based on the Experience Risk Minimization principle and theβ-stable uniform algorithm, proposed a edge weight statistical properties based merge criterion which do well in keeping the integrity of the region and avoid over segmentation. It applies to the high-resolution remote sensing image segmentation.3. Studied and realized a MST based object-oriented multilevel and multi-scale irregular pyramid high resolution image segmentation method. A Disjoint-set data structure which can retain the spatial adjacency is designed. The kruskal MST algorithm and simple graph is used to realize the multilevel and multi-scale image segmentation and to descript the segment results. It provided the technical support for multi-scale image interpretation and object recognition.4. A region statistical properties based multilevel image segmentation criterion is designed. After initial segmentation, take the regions as vertices, adjacent regions are linked by edges to construct new level graph model. Using the region statistical properties, histogram distance is used to calculate the edge weight. The new merge criterion is used to the MST based higher level image segmentation. This method can ensure integrity of objects.The QuickBird high resolution multispectral images and panchromatic images are used to test the segmentation method. The analysis and assessment are given in this paper. The knowledge based object recognition experiments proved the segmentation method proposed in this paper is useful.
|
PDF Full Text Request