Contour Segment Feature Weighting in Infrared Automatic Target Recognition

We study automatic target recognition (ATR) in infrared (IR) imagery by applying two recent computer vision techniques, Hierarchical description of contour and Bag-of-Words (BoW). We propose the idea of contour segment (CS) features which are extracted from the outer contour of IR target image and we developed a new weighted scalable vocabulary tree (SVT) that is learned from a set of training images to support efficient and scalable Bow-based ATR. We develop a weighted BoW model to improve the ATR performance by the feature weighting. Different from traditional BoW model, the saliency information is incorporated into feature weighting to enhance the voting confidance in Bow-based classification. The proposed CS features weighting ATR algorithm is evaluated against recent relevance grouping of vocabulary (RGV) approaches that reportedly outperform traditional methods. Experimental results on dedicated IR dataset demonstrate the advantages of the newly proposed algorithm outperform the baseline BoW method and the recent RGV approach.