Edge Grouping with a Novel Shape Model for Object Detection
Lei Ma, Bitao Jiang · 2013
This paper presents a new method for object detection by edge grouping. This method can detect the boundaries of objects under complex background where the object contours are partly occluded or missing during contour extraction. Our method is adapted to detect the objects with not only closed boundaries but also open-boundaries. There are three contributions in this work. First, the shape of an object is represented by a novel Turn Angle Probabilistic Sequence Model (TAPSM) which originates from HMM. This shape model is robust for noisy images. Second, edge grouping is defined in a sequential search procedure based on TAPSM, which reduces the search complexity. Third, a linear discrimination weighted by probabilities is developed to evaluate the similarity between the detected line sequences and the model. We employ this approach to the problem of detecting warships from satellite imagery, and experimental results demonstrate the high performance of the proposed method.