Object-Oriented Multi-Scale Segmentation Of High Resolution Remote Sensing Image

Remote sensing image is a kind of high-resolution image. It has the characteristics of rich detail information, clear geometric structure and important textural information. Traditional remote sensing image processing methods were mainly pixels-oriented and didn't take the image shape features and object topological structure into account, which caused the difficulties in qualitative analysis. Nowadays, object-oriented processing approaches are brought up, which integrate spectrum characters, textural and shape information. In object-oriented remote sensing image processing methods, a region is processed as an object. In traditional pixel-oriented processing methods, the processing unit is a pixel. Object-oriented processing methods can analyze images from both pixel level and region level which greatly use the image geometric and structure information.The most important procedure in object-oriented processing methods is image segmentation. The quality of segmentation greatly impacts the precision of the following analysis, such as classification, recognition and comprehension. But because of the complexity of remote sensing image and the difficulties existed in image segmentation techniques, the precise segmentation of remote sensing image has been concerned as an intractable problem. One of the most important reasons is that only one scale was used to segment remote sensing images. Using multi-scale segmentation methods can solve the above problems by setting different scales according to the different regions.The Graph theory now plays a important part in the image segmentation. It first builds the mapping relation of image and graph, then sets an energy function of the graph according to the remarkably weights, the energy functions can be transformed into a generalized eigenvalue equation, we can solve the equation to get the minimum eigenvector which will lead to the result of segmentation, at last, we mapped the graph segmentation result back to image. Because the construction of graph and the extraction of remarkable weights can be based both on pixel and on object, the Graph theory based methods are good object-oriented image segmentation methods.In this paper, we present a new method which uses the spectrum, shape and texture features. It is the improved weighted aggregation algorithm based on the N-cut theory, from the pixel level to object level, coarse level to fine level. First, the improved weighted aggregation algorithm builds the graph according to the image, then considers the spectrum, shape and texture features to compute the vertex weights, after that, it aggregates the regions according to N-cut to start the segmentation. During the process, we need to update the weights. When implementing the iterative segmentation, we can define the appropriate scale eigenvalue equation according to different regions, and that is multi-scale segmentation.The experiments demonstrate that our methods have good segmentation results, and the extracted contours are consisted with the real edge of objects in the image.