Skip to Main content Skip to Navigation

Détection et classification de décors gravés sur des céramiques anciennes par analyse d’images

Abstract : The ARCADIA project aims to develop an automatic method for analyzing engraved decorations on ceramic sherds to facilitate the interpretation of this archaeological heritage. It is to replace the manual and tedious procedure carried out by the archaeologist since the corpus increased to more 38000 sherds. The ultimate goal is to grouping all the decorations created with the same wheel by a poter. We developped a complete chain from the 3Dscanning of the sherd to the automatic classification of the decorations according to their style (diamonds, square, chevrons, oves, etc). In this context, several contributions are proposed implementing methods of image analysis and machine learning. From the 3Dpoint cloud, a depth map is extracted and an original method is applied to automatically detect the salient region centered onto the decoration. Then, a new descriptor, called Blob-SIFT, is proposed to collect signatures only in the relevant areas and characterize the decoration to perform the classification. This approach adapted to each sherd, allows both to reduce significantly the mass of data and improve classification rates. We also use deep learning, and propose an hybrid approach combining local features extracted by Blob-SIFT with global features provided by deep learning to increase the classification performance.
Complete list of metadatas

Cited literature [116 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Thursday, April 11, 2019 - 9:33:08 AM
Last modification on : Wednesday, November 20, 2019 - 1:42:54 AM


Version validated by the jury (STAR)


  • HAL Id : tel-02096056, version 1


Teddy Debroutelle. Détection et classification de décors gravés sur des céramiques anciennes par analyse d’images. Traitement du signal et de l'image [eess.SP]. Université d'Orléans, 2018. Français. ⟨NNT : 2018ORLE2015⟩. ⟨tel-02096056⟩



Record views


Files downloads