Affine Transform Reconstruction Based Target Recognition

Automatically target recognizing must concern the influence of imaging environment changing, such as the difference of imaging situation, imaging time and so on. Affine invariance is an important research field about it.In usual ATR systems, feature matching is used to recognize and classify targets after invariant vectors have been extracted. Because of the information losing during the process of extracting feature vectors, this method has some limitation.In this paper, a novel ATR system framework is proposed. Unlike the case of usual framework, the new strategy is: the invariant vectors are only used to object registration. First the feature matching is used to search targets relationship between test and reference image, and then, depending on this relationship, reference image is recovered to the same coordinate system of the test image by an affine transformation. After this image recovering, various target recognizing method——such as image matching——could be used.The key technique of this novel framework is to design an effective affine invariant vector, which has the capability of object registration and is unaffected by image noise, detail changing and target aberrance. Skelecton convex hull, abbreviated as SCH, is introduced in this paper. SCH is the convex hull of the object's geometric structure, so it has the stability of insensitive to noise and detail changing. At the same time it has the local controllability inherited from convex hull. So invariant feature matching based on SCH could register targets in vile environment.SCH requests that object's skelecton is stable under various affine transform. Current morphological thinning algorithm usually produces skeleton with too much offshoot, and in different transformation coordinate system, skeleton of an object changes too much. In this paper, a burr removing iteration algorithm is designed. As a posterior disposal of morphological thinning processing, this algorithm could effectively make the skeleton expressing object's geometric structure more correctly and steadily to satisfy the requests of target invariant apperception.Experiments proved that the novel ATR method could recognize targets correctly in spite of the difference of affine transform, target detail changing and target aberrance.