Method for the selection of attributes for statistical learning for object detection and recognition

Attribute selection method for statistical learning of descriptors intended to enable automatic recognition and/or automatic detection of an object from a set of images, method characterized by the following steps:

• obtain a mask of the object in each image containing said object to be recognised,
• define at least one set of descriptors using their geometric shape and/or apparent specific physical characteristics,
• select at least one set of descriptors as a function of their geometric shape and/or apparent specific physical characteristics,
• calculate attributes associated with these descriptors and said specific physical characteristics,
• for each descriptor and for each image, define a semantic conformity score with the mask of the object to be recognised in the image previously calculated representing the conformity level of the geometric shape of said descriptor with the mask of the object to be recognised in the image,
• sort the descriptors as a function of their corresponding scores,
• select descriptors with the highest scores to perform said statistical learning,
• measure a statistical property on a combination of adjacent geometric shapes and non-adjacent geometric shapes, using said descriptors,
• focus on the most specific zones in each of the classes as iterations in the learning phase continue, in order to eliminate background zones (12) at the beginning of the processing and to discriminate between classes at the end of the processing.

Attribute selection device for making statistical learning of descriptors intended to enable automatic recognition and/or detection of an object from a set of images, characterized in that it comprises:

• means for segmenting each image containing said object to be recognised,
• means for defining at least one set of descriptors as a function of their geometric shape and/or apparent specific physical characteristics,
• means for calculating attributes associated with these descriptors and said specific physical characteristics,
• means for selecting at least one set of descriptors as a function of their shape and attributes calculated on this shape,
• means for defining, for each descriptor and for each image, a semantic conformity score with the mask of the object to be recognised in the image representing the conformity level of the geometric shape of said descriptor with the mask of the object to be recognised in the image,
• means for sorting descriptors as a function of their respective scores,
• means for selecting descriptors with the highest scores to perform said statistical learning, and to measure a statistical property on a combination of adjacent geometric shapes and non-adjacent geometric shapes.