Segmentation for remote sensing image with shape and spectrum prior

Segmentation of objects with a high accuracy is the key step to achieve automatic interpretation and classification of remote sensing images. However, degradation caused by turbulent motion of the atmosphere, blur due to cloud and disturbance of light will all smear the images, the most vigorously studied active contour model still grapples hard with weak edges, low contrast and partial occlusions. To remedy these drawbacks, a variational segmentation method with constraints of shape and spectrum prior is proposed. The shape prior energy term is defined to ensure the similarity between shape prior and the evolving curve. The spectrum prior energy term is put forward to define the speed of the evolving curve. Kullback-Leibler distance is adopted to measure the spectrum similarity between the spectrum signature of the object and the spectrum prior. Finally, the prior knowledge is incorporated into the variational framework and the energy minimization is implemented by the gradient descend flow. The experimental results show that this approach achieves a higher accuracy, in comparison with the representative data-driven and recently proposed shape-driven active contour models.