Shape model-based region grouping, a method for deformable object detection and retrieval

A region grouping and splitting paradigm for deformable shape detection and recognition is proposed. In the proposed paradigm, region-based feature and boundary information are combined with a deformable shape model. A statistical shape model is introduced to enforce the prior probabilities on global, parametric deformations for each object class. The approach can effectively handle cluttered images of multiple objects and images with shadows. The system avoids merging objects with the background, shadows, and adjacent objects. During segmentation, the system also recovers a shape model for each detected object. Segmentation and identification are executed simultaneously. The recovered shape models can be used in object recognition and content-based image retrieval.; The shape model employed is a 2D region-based model. Statistical methods are used in training the model from examples and during object recognition using the recovered shape parameters. The shape deformation formulation used is a combination of affine transformations and bending deformations, but other kinds of deformations can be easily implemented and employed in the algorithm. After the model training stage, the system is fully automatic in object detection and shape description.; The object detection algorithm includes three basic stages: an over-segmentation stage, a model-based region merging stage, and a model-based region splitting stage. An index tree-based model fitting method is used to accelerate the model fitting procedure. The method utilizes hierarchical clustering for coarse searching and a neural network for finer searching. Use of the index tree results in a speedup of about one order of magnitude compared to fitting with the downhill-simplex method.; Synthesized images, images taken under controlled laboratory conditions, pictures from the world wide web, and medical images from microscope are used for the experiments that test the new method's segmentation accuracy and efficiency. An image database retrieval system based on the object detection system has been implemented and tested, and results are promising.; Although in the experiments, color features are used in segmentation, the underlying formulation is more general. The algorithm can be used to detect regions of interest based on any image homogeneity predicate; e.g., texture, color, or motion.