The Class-Specific Down-Looking Target Localization Combining Recognition and Segmentation

In the complex down-looking background, it is difficult to accurately localize various targets because of target deformation and background clutter. In this paper, we develop a target detection algorithm that incorporates bottom-up target segmentation and top-down target recognition. There are two main steps in the algorithm: hypotheses generation (top-down) and hypotheses verification (bottom-up). In the generation step, the study makes an improvement on shape feature, which is more robustness to target deformation. The improved shape feature is used to generate the hypotheses of target locations and figure-ground masks. In the hypotheses verification step, the study firstly computes feasible target segmentation that is consistent with top-down target hypotheses. And then a false positive pruning procedure is proposed. The study also finds the fact that the pruned false positive regions do not align with target segmentation for many down-looking targets. The experimental tasks demonstrate that the algorithm can be high precision and recall with a few positive target-training images and that the algorithm, and be generalized to many target classes.