fr | en
Laboratoire Angevin de Recherche en Ingénierie des Systèmes
  • Index
  • >News
  • >Defense of Mr. Jérémy CHOPIN's thesis

Separated by coma

Defense of Mr. Jérémy CHOPIN's thesis9:30 am | POLYTECH ANGERS | AMPHI E | 62, avenue Notre-Dame du Lac 49000 ANGERS

Subject : Deep Learning and Structural Knowledge for Image Analysis.

Director of thesis : Mr. Jean-Baptiste FASQUEL

Abstract

Semantic image segmentation is a key task in computer vision for many applications. In this context, deep learning has made remarkable progress, but sometimes leads to gross errors and requires a lot of representative data for training. In this context, our work focuses on providing "high-level" structural information, which corresponds to the relationships between different regions of a scene, and which can be easily extracted from a few annotated images or formulated by an expert (e.g. anatomical knowledge such as "the heart is between the lung" for medical image analysis). We designed a method, integrating this information as output from a deep neural network, by means of inexact graph matching formulated as a two-stage quadratic assignment problem (QAP). To cope with the combinatorial nature of the QAP, we also incorporate a method for sequential matching of nodes in a learned order by reinforcement learning. This method is evaluated on two public databases (face recognition and brain imaging) and with different neural network architectures. The results show that it is possible to significantly increase the performance of deep learning, especially when little training data is available.

Scroll