Shape constrained discrete dynamic contours for noisy object segmentation

In this paper, we focus on using shape information of a known class of objects to match an active contour model (ACM) with the boundary of an object in a noisy scene. The problem is addressed as finding two similar shapes in the presence of noise, where the edges of the desired object are not clearly distinguishable, during the iterative process of finding the energy-minimized active contour. The shape information is obtained through the transformation, scale, and rotation invariant Fourier descriptors (FD). During a training phase, the principal component analysis (PCA) of the FD is performed to find the modes of the variation of the FD. This is different from active shape models (ASM) that deal with shape through a point distribution model (PDM) in the spatial domain. Our proposed method overcomes the difficulties associated with landmark localization and shape normalization in ASM. We replace the shape constraint with the internal energy of ACM. The contour is constrained to an allowable space in FD domain (shape information) and cannot freely deform according to the external energy. Experimental results show a faster convergence in comparison to the original ACM and less user interaction in the training phase compared to ASM. Also, the resulting contours are more similar to the mean of the expert's manual segmentation.