Adaptive outlier elimination in image registration using genetic programming

In feature-based methods, outlier removal plays an important role in attaining a reasonable accuracy for image registration. In this paper, we propose a genetic programming (GP) based adaptive method for outlier removal. First, features are extracted through the scale-invariant feature transform (SIFT) from the reference and sensed images which were initially matched using Euclidean distance. The classification of feature points into inliers and outliers is done in two stages. In the first stage, feature vectors are computed using various distance and angle information. Feature points are categorized into three groups; inliers, outliers and non-classified feature (NCF) points. In the second stage, a GP-based classifier is developed to classify NCF points into inliers and outliers. The GP-based function takes features as an input feature vector and provides a scalar output by combining features with arithmetic operations. Finally, registration is done by eliminating the outliers. The effectiveness of the proposed outlier removal method is analyzed through the classification and positional accuracy. The experimental results show a considerable improvement in the registration accuracy.