Clustering based content and color adaptive tone mapping

By extracting image luminance channel and separating it into a base layer and a detail layer, the Retinex theory has been widely adopted for tone mapping to visualize high dynamic range (HDR) images on low dynamic range display devices. Many edge-preservation filtering techniques have been proposed to approximate the base layer for Retinex image decomposition; however, the associated tone mapping methods are prone to halo artifacts and false colors because filtering methods are limited in adapting the complex image local structures. We present a statistical clustering based tone mapping method which can more faithfully adapt image local content and colors. We decompose each color patch of the HDR image into three components, patch mean, color variation and color structure, and cluster the patches into a number of clusters. For each cluster, an adaptive subspace can be easily learned by principal component analysis, via which the patches are transformed into a more compact domain for effective tone mapping. Comparing with the popular edge-preservation filtering methods, the proposed clustering based method can better adapt to image local structures and colors by exploiting the image global redundancy. Our experimental results demonstrate that it can produce high-quality image with well-preserved local contrast and vivid color appearance. Furthermore, the proposed method can be extended to multi-scale for more faithful texture preservation, and off-line subspace learning for efficient implementation.