TY - JOUR
T1 - 2-D object recognition by indexing through a modified ART-2 neural network
AU - Sossa, J. Humberto
AU - Villela, P. Rayón
AU - Nazuno, J. Figueroa
PY - 1998
Y1 - 1998
N2 - A technique for the recognition of possibly occluded planar objects from a single image in the presence of projective deformations when the camera-plane is not perpendicular to the object-plane is presented. The technique is divided into two phases. During the first phase (database construction), an object is first decomposed into a set primitives called metasegments. These are groups of consecutive segments obtained from the corners of the object's contour coded by three geometric invariants: the type, and the two four and the five point-dependent affine/projective invariants. Each resulting code is then used to build the corresponding database of models. It is composed of a standard ART-2 NN connected to a Memory Map (MM), a set of logical AND gates, an evidence-register (an adder) and a set of comparators. The ART-2 NN has as input the code of a metasegment and a number of outputs equal to the number of different metasegments trained to the NN. The Memory Map has as many rows as outputs provided by the ART-2 NN and as many columns as objects used to train the NN. Each of the MM's locations contains a value. This value represents the number of metasegments present in each trained object. The indexing phase is divided in two stages: candidate selection and candidate reduction. During candidate selection, an image containing one or more possibly occluded objects is first preprocessed to obtain the corresponding contours. Each contour is then decomposed into its metasegments and coded as explained above. Each code is next used by the trained ART-2 NN to retrieve from the previously constructed Memory Map the list of objects that had possibly produced the corresponding metasegment. At the end of this process, we will have in an evidence-register the number of times an object was voted for during candidate selection. A selection-threshold is finally used, during the candidate reduction stage, to select those objects most possibly present in the test image. The system's performance is tested with a set of polygonal objects.
AB - A technique for the recognition of possibly occluded planar objects from a single image in the presence of projective deformations when the camera-plane is not perpendicular to the object-plane is presented. The technique is divided into two phases. During the first phase (database construction), an object is first decomposed into a set primitives called metasegments. These are groups of consecutive segments obtained from the corners of the object's contour coded by three geometric invariants: the type, and the two four and the five point-dependent affine/projective invariants. Each resulting code is then used to build the corresponding database of models. It is composed of a standard ART-2 NN connected to a Memory Map (MM), a set of logical AND gates, an evidence-register (an adder) and a set of comparators. The ART-2 NN has as input the code of a metasegment and a number of outputs equal to the number of different metasegments trained to the NN. The Memory Map has as many rows as outputs provided by the ART-2 NN and as many columns as objects used to train the NN. Each of the MM's locations contains a value. This value represents the number of metasegments present in each trained object. The indexing phase is divided in two stages: candidate selection and candidate reduction. During candidate selection, an image containing one or more possibly occluded objects is first preprocessed to obtain the corresponding contours. Each contour is then decomposed into its metasegments and coded as explained above. Each code is next used by the trained ART-2 NN to retrieve from the previously constructed Memory Map the list of objects that had possibly produced the corresponding metasegment. At the end of this process, we will have in an evidence-register the number of times an object was voted for during candidate selection. A selection-threshold is finally used, during the candidate reduction stage, to select those objects most possibly present in the test image. The system's performance is tested with a set of polygonal objects.
UR - http://www.scopus.com/inward/record.url?scp=0041595417&partnerID=8YFLogxK
U2 - 10.1016/s0957-4174(97)00061-4
DO - 10.1016/s0957-4174(97)00061-4
M3 - Artículo
SN - 0957-4174
VL - 14
SP - 199
EP - 210
JO - Expert Systems with Applications
JF - Expert Systems with Applications
IS - 1-2
ER -