cahier jupyter<\/a> offre une demonstration pratique de l’usage des mod\u00e8les YOLOv5 pr\u00e9-entra\u00een\u00e9s pour segmenter des commentaires disponibles dans l’Internet Archive.<\/p>\n","protected":false},"excerpt":{"rendered":"L’analyse de la mise en page est une \u00e9tape charni\u00e8re du traitement de documents. La t\u00e2che consiste \u00e0 segmenter une page en r\u00e9gions puis \u00e0 classifier ces r\u00e9gions selon une … <\/p>\n","protected":false},"author":1002569,"featured_media":0,"template":"","meta":{"_seopress_titles_title":"","_seopress_titles_desc":"","_seopress_robots_index":"","_seopress_robots_follow":"","_seopress_robots_imageindex":"","_seopress_robots_snippet":"","_seopress_robots_primary_cat":"","_seopress_robots_breadcrumbs":"","_seopress_robots_freeze_modified_date":"","_seopress_robots_custom_modified_date":"","_seopress_robots_canonical":"","_seopress_social_fb_title":"","_seopress_social_fb_desc":"","_seopress_social_fb_img":"","_seopress_social_fb_img_attachment_id":0,"_seopress_social_fb_img_width":0,"_seopress_social_fb_img_height":0,"_seopress_social_twitter_title":"","_seopress_social_twitter_desc":"","_seopress_social_twitter_img":"","_seopress_social_twitter_img_attachment_id":0,"_seopress_social_twitter_img_width":0,"_seopress_social_twitter_img_height":0,"_seopress_redirections_value":"","_seopress_redirections_enabled":"","_seopress_redirections_enabled_regex":"","_seopress_redirections_logged_status":"","_seopress_redirections_param":"","_seopress_redirections_type":0,"_seopress_analysis_target_kw":"","_seopress_news_disabled":"","_seopress_video_disabled":"","_seopress_video":[],"_seopress_pro_schemas_manual":[],"_seopress_pro_rich_snippets_disable_all":"","_seopress_pro_rich_snippets_disable":[],"_seopress_pro_schemas":[],"_tribe_events_status":"","_tribe_events_status_reason":"","footnotes":""},"tags":[],"tribe_events_cat":[],"class_list":["post-868","tribe_events","type-tribe_events","status-publish"],"_links":{"self":[{"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/tribe_events\/868","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/tribe_events"}],"about":[{"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/types\/tribe_events"}],"author":[{"embeddable":true,"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/users\/1002569"}],"version-history":[{"count":3,"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/tribe_events\/868\/revisions"}],"predecessor-version":[{"id":882,"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/tribe_events\/868\/revisions\/882"}],"wp:attachment":[{"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/media?parent=868"}],"wp:term":[{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/tags?post=868"},{"taxonomy":"tribe_events_cat","embeddable":true,"href":"https:\/\/wp.unil.ch\/asanumerica\/wp-json\/wp\/v2\/tribe_events_cat?post=868"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
![\"\"](\"https:\/\/wp.unil.ch\/asanumerica\/files\/2023\/02\/ajmc-layout-regions-238x300.png\")
L’analyse de la mise en page est une \u00e9tape charni\u00e8re du traitement de documents. La t\u00e2che consiste \u00e0 segmenter une page en r\u00e9gions puis \u00e0 classifier ces r\u00e9gions selon une taxonomie donn\u00e9e. Si les mod\u00e8les \u00e9tat de l\u2019art affichent des r\u00e9sultats concluants sur des documents g\u00e9n\u00e9riques (livres, cartes, relev\u00e9s comptables\u2026), ils sont en revanche mis au d\u00e9fi par la complexit\u00e9 graphique et le multilinguisme des commentaires classiques. D\u2019une \u00e9dition \u00e0 l\u2019autre, composition de ces derniers peut varier consid\u00e9rablement et les r\u00e9gions les plus importantes y sont mieux caract\u00e9ris\u00e9es par leur contenu s\u00e9mantique que par leur apparence visuelle. Ce contexte particulier appelle une comparaison entre des approches textuelles, visuelles et hybrides. Nous \u00e9valuons donc les performances de deux transformers (LayoutLMv3 et RoBERTa) et d’un r\u00e9seau de d\u00e9tection d’objets (YOLOv5). Si nos exp\u00e9riences donnent un un net avantage \u00e0 YOLOv5, nous listons \u00e9galement une s\u00e9rie mises en garde \u00e0 l\u2019\u00e9gard de ces premiers r\u00e9sultats.<\/p>\n