A bio-inspired quaternion local phase CNN layer with contrast invariance and linear sensitivity to rotation angles

Deep learning models have been particularly successful with image recognition using Convolutional Neural Networks (CNN). However, the learning of a contrast invariance and rotation equivariance response may fail even with very deep CNNs or by large data augmentations in training. We were inspired by...

Full description

Bibliographic Details
Authors: Moya Sánchez, Eduardo Ulises, Xambó Descamps, Sebastián|||0000-0001-5056-9818, Sánchez Pérez, Abraham, Salazar Colores, Sebastián, Martínez Ortega, Jorge, Cortés García, Claudio Ulises|||0000-0003-0192-3096
Format: article
Publication Date:2020
Country:España
Institution:Universitat Politècnica de Catalunya (UPC)
Repository:UPCommons. Portal del coneixement obert de la UPC
Language:English
OAI Identifier:oai:upcommons.upc.edu:2117/422915
Online Access:https://hdl.handle.net/2117/422915
https://dx.doi.org/10.1016/j.patrec.2019.12.001
Access Level:Open access
Keyword:Convolution
Deep learning
Image recognition
Rotation
Multilayer neural networks
Mammals
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació::Processament del senyal::Processament de la imatge i del senyal vídeo
Description
Summary:Deep learning models have been particularly successful with image recognition using Convolutional Neural Networks (CNN). However, the learning of a contrast invariance and rotation equivariance response may fail even with very deep CNNs or by large data augmentations in training. We were inspired by the V1 visual features of the mammalian visual system to emulate as much as possible the early visual system and add more invariant capacities to the CNN. We present a new quaternion local phase convolutional neural network layer encoding three local phases. We present two experimental setups: An image classification task with four contrast levels, and a linear regression task that predicts the rotation angle of an image. In sum, we obtain new patterns and feature representations for deep learning, which capture illumination invariance and a linear response to rotation angles.