This paper addresses the issue of retrieving images based on visual content, according a particular attention to the semantic dimension of information retrieval. A brief review of existing Image Retrieval Systems is provided, hilighting a major drawback of these prototypes, namely the lack of integration between classical ``semantic search'', and visual similarity retrieval (i.e. content-based retrieval). A new approach is proposed, that tries to integrate a real ``semantic'' dimension into visual content-based image retrieval. Images may be retrieved either by keyword matching (for manually indexed images), or by visual content matching. This approach is then particularly adapted to partially indexed databases, and may be used in a semi-automatic indexing tool.