METHOD FOR AUTOMATIC IMAGE ANNOTATION

Abstract

METHOD FOR AUTOMATIC IMAGE ANNOTATION. A method for correcting annotations in an image database containing several thousand images and several thousand associated annotations is described. In a series of steps prior to the actual method, operating means are used in the form of a known electronic neural network, for example, a convolutional neural network capable of classifying objects or parts of the human body in these images. Then, the result provided by the neural network is used to detect discrepancies between the annotations initially present in the database and the results of the neural network.When such a discrepancy is detected and the neural network's detection accuracy exceeds a PS threshold, the annotated database is updated by supplementing the annotation data with the additional data provided by the neural network. The process is then iterated, returning to the first step but with the new set of annotations. Finally, when the processing methods used no longer detect any discrepancies, the process is complete and the new annotated database can be used.

Images