SimDSR: Simultaneous detection and segmentation for repetitive patterns

Abstract

Detection and segmentation are two strongly correlated tasks, yet typically solved separately with different techniques for repetitive patterns or simultaneously solved for generic patterns. We propose a Simultaneous Detection and Segmentation model for Repetitive pattern analysis (SimDSR) in real-world images. The proposed SimDSR model implements a joint patch-contour alignment between the multi-scale over-segments, where the optimal repetitive segment is acquired by minimizing the stable alignment feedbacks. The novel joint alignment is exploited to measure the parametric warp discrepancy with the 2-D affine Lie group on pixel intensity and the Gaussian mixture L2 distance on boundaries. Thus, a repetitive segment rather than a repetitive patch is predicted gradually growing into a series of repetitions. Extensive experiments and comparative analysis have demonstrated encouraging performance of the proposed algorithm for both detection and segmentation of repetitive patterns.