Computerimplementiertes Verfahren zum Trainieren eines Vorwärtsgekoppelten Tiefenlernmodells zur Detektion und Instanzsegmentierung von Objekten

Details

Veröffentlichungs-Nr.

EP4776243

Anmeldetag

14. Januar 2025

Veröffentlichung

15. Juli 2026

Rechtsraum

EP

Abstract

Computer implemented method for training a feed-forward deep learning model intended to the detection and instance segmentation of objects, comprising the steps of: inserting (S1) a gradient reversal layer (GRL) after a hidden layer of the deep learning model, inserting (S2) a domain classifier layer (DCL) after the gradient reversal layer (GRL); introducing (S3) a set of input values (I) of training data into an input layer of the deep learning model; obtaining (S4) a set of output values (O) in response to said input values (I); calculating (S5) a loss function of the last hidden layer of the deep learning model, calculating (S6) the gradient of the loss function; propagating backwards the loss function and repeating steps (S5) and (S6) for each hidden layers and adjusting (S7) the weight of each hidden layer, so as to minimize the gradient of the loss function.

Anmelder

Firma

Airbus Defence and Space, S.A.U.

Land

🇪🇸 Spanien