Skip to Main content Skip to Navigation

Approches connexionnistes pour la vision par ordinateur embarquée

Abstract : To design embedded computer vision systems, two axes can be considered. The first focuses on designing new, more powerful, digital devices that can efficiently implement complex algorithms. The second targets the development of new, lightweight computer vision algorithms that can be effectively implemented on digital embedded systems. In this work, we favor the second axis by using connectionist models. In this context, we focus on two models of artificial neural networks: cluster-based networks and convolutional networks. The first model we use, i.e. cluster-based network, was never been used to perform computer vision tasks before. However, it seemed to be a good candidate to design embedded systems, especially through dedicated hardware architectures implementation. The goal was first to find out the kinds of tasks that could be performed using this network model. This model has been designed to implement associative memories which can come close to problems such as content- based image retrieval in computer vision domain. This type of application massively uses approximated nearest neighbor search algorithms which makes it a good candidate to focus on. The second type of network studied in this work, called convolutional network, is very popular to design computer vision systems. Our goal here was to find different ways to simplify their complexity while maintaining high performance. In particular, we proposed a technique that involves re-training quantified networks.
Document type :
Complete list of metadatas

Cited literature [61 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Wednesday, December 11, 2019 - 3:16:08 PM
Last modification on : Friday, September 25, 2020 - 3:35:51 AM
Long-term archiving on: : Thursday, March 12, 2020 - 6:28:51 PM


Version validated by the jury (STAR)


  • HAL Id : tel-02404823, version 1


Robin Danilo. Approches connexionnistes pour la vision par ordinateur embarquée. Autre. Université de Bretagne Sud, 2018. Français. ⟨NNT : 2018LORIS518⟩. ⟨tel-02404823⟩



Record views


Files downloads