| Cultural institutions around the world have started over the past two decades a safeguarding policy of their collections. Their goals are the preservation of digital works, the monitoring of their condition, and the improvement of the accessibility to public and professionals. This last goal is very important for museums because it is an integral part of their mission. This policy has led museums to build extremely rich databases (often more than 10000 images) and available on the Internet.;Museums mostly use metadata to index their databases, i.e. text fields handfilled by an expert. This task is valid and useful for factual data such as the artist's name or the date of creation, though very expensive. But for every information that is related to the content of the work (color palette, style, represented subjects), this policy is subject to too much subjectivity and non reproducibility.;The main objective of this thesis is to propose new descriptors and new similarity measures adapted to the specific content of artworks. Most of the approaches proposed in the scientific literature are based on statistical descriptors characterizing certain aspects of local pictorial content. To better understand the specific characteristics of pictorial content, we inspired ourselves with literature dealing with the study and analysis of the artistic content of a work in art history. Several historians, psychologists and experts of cognition have indeed addressed this medium to understand the mechanisms of the artist vision and observer vision. A work of 2D art has indeed the peculiarity of being visually produced through a human visual system.;This study has led us to two different approaches proposed in this thesis. These approaches offer to consider two complementary aspects of the pictorial content that can meet different search criteria. The first part focuses on the study of the spatial organization of colors in a work. This feature describes the spatial arrangement of homogenous regions. This feature is very similar to what we often call the pictorial composition of an artwork. The way the regions are organizing themselves in a colorful work is a descriptor that applies to all types of artworks (paintings, photos, drawings and so on). It is also linked to the concept of balance. The composition of a painting can influence the visual impression regardless of the semantic content of the image. This organization is often at the heart of the creative process of the artwork. The methodology introduced in this part proposes to model the comparison between two images according to the spatial organization of colors by using an optimal transport problem. We also propose an adaptive threshold on the distance based on a statistical approach. To evaluate the performance of these methods, we present in this thesis many results on five different bases gathering approximately 60000 images.;Being able to view and browse such voluminous databases is a complex task. Artworks are unique compared to natural images because they are often stylized. This property induces a subjectivity which can lead to misunderstanding of the scene (through multiple ambiguities), and also a specific visual impression delivered to the viewer when he looks at the image. We do not look at a painting with the same attention as at a natural image. Visitors in a museum can expect a certain aesthetic experience, abstract and ephemeral. This content, and this particular use of the works require specific search criteria that are complementary to classical content-based criteria.;The second part of this thesis focuses on the study of the geometrical content in line-drawings. Particularly suited to graphic arts, the proposed methodology introduces a possible description of pictorial content related to the lines drawn by the artist and the visual impact transmitted through them. This type of content is close to the pictorial style of the artist but it is also inseparable from the represented subject. This methodology is based firstly on an unsupervised extraction of contour lines. These contours are extracted from the map of topological levels lines in the image. We propose in this thesis a new method to filter the tree of levels lines and obtain these contours. We then also propose a set of methods that analyze these geometrical contours, estimating inflections points, junctions of lines, endpoints of lines, corners and line thickness. A vector of 11 descriptors is computed using these geometrical components. We compare this indexing method with a recent method based on the curvature scale space (CSS). The results obtained on a classified database of 105 drawings and on a unclassified database of 259 drawings is used to evaluate the performance of the methodology. |
