Abstract: Automatic recognition of partially occluded objectsthat are sensed by imaging sensors is a challengingproblem in image understanding (IU), automatic targetrecognition (ATR), and computer vision fields. In thispaper I address this problem by using a geneticalgorithm (GA) as part of a model-based recognitionscheme. The partially occluded object segments arerotated, translated, and scaled. Then each transformparameter is encoded into a binary string and used in agenetic algorithm. The suggested transformation is thenapplied to the sensed segment and the resulting objectis matched against a library of stored targets. Thefitness criterion is a distance function that measuresthe similarity between the segmented object and thestored target models. The GA by performing the processof mutation, reproduction, and crossover suggestsoptimum transform parameter sets. The empirical resultsof the application of the approach on a set of realladar data of military targets shows that correctrecognition for up to 50% target occlusion is possible.!7
展开▼