Automatic color patch selection for painting identification

A new method to define a digital signature of a painting is presented. This signature is composed by a set of patches that are automatically selected as the images regions containing the most relevant information. The region selection is applied on a combination of saliency maps related to different features concerning intensity, color and visual saliency. We present the generation of the feature saliency map exploiting low-level feature representations, and the new algorithm for selecting the most relevant regions. The position and actual size of these regions is not a-priori fixed but is function of the saliency maps, i.e. of the painting content. We present the experimental results on several painting, discussing the trade-offs among image features parameters values and the selected regions. Introduction Many projects are carried on accurate paintings digital acquisition purposes [1] [2]. We present a method working on the final color images that performs selection of the most interesting regions of a painting in order to further make it possible to identify and to characterize a painting by its content. This work aims to perform an automatic selection of the interesting regions, which has to be adaptive to the relative richness of the image content. The signature of a painting is defined as salient parts related to the image content itself (radiometric signal) and to the human visual perception. These salient parts are extracted from a combination of feature maps (Figure 1). Figure 1. General principle for the selection of a salient region The construction of these maps is firstly presented, with a description of the features and their representations selected to characterize the richness of a painting. The method for the selection of the salient regions from these features is then exposed and results are presented and analyzed. Image description Being an automatic area selection, the method uses features in a large field of application, in order to cover the large set of existing paintings styles and techniques. The patch selection is applied on a combination of saliency maps related to different features concerning intensity, color, texture and visual saliency. The general approach to compute these maps is illustrated in Figure 2. A spatial “saliency” map is calculated for each feature representation, with neighbour consideration. Each pixel of the image is associated a 37-dimensional feature vector composed by: a 7-dimensional feature vector related to intensity, a 26dimensional feature vector related to color (RVB, rvb, HS, CbCr), and a 4-dimensional feature vector related to visual saliency. Figure 2. General principle of feature map computation: for each pixel position, a neighbour is defined, the feature is computed using the neighbour influence, leading to one pixel value in the final feature map