Fast Symmetry Detection with Deep Learning and GeConv

LA_{TEX TikZposter}

Fast Symmetry Detection with

Deep Learning and GeConv

PhD Candidate: T. Mkhoyan Department: ASM

Supervisor 1st: Dr.ir.R. De Breuker Supervisor 2nd: Dr.ir.C.Visser

Promotor Dr.ir.R. De Breuker Contact: T.mkhoyan@tudelft.nl

Fast Symmetry Detection with

Deep Learning and GeConv

PhD Candidate: T. Mkhoyan Department: ASM

Supervisor 1st: Dr.ir.R. De Breuker Supervisor 2nd: Dr.ir.C.Visser

Promotor Dr.ir.R. De Breuker Contact: T.mkhoyan@tudelft.nl

Aim

This study: Develop Fast 2-axis reflectional symmetry de-tection routine for estimation of aircraft wing orientation. Two methods developed and

com-pared: traditional computer

vision (GeConv ) versus pure Deep learning (RotNet ).

High level: Robust machine learning pipeline for prediction of wing deflection for aeroservoe-lastic control from raw images.

The Experiment

Markers

Gust Wing oscillations

ima_ge u Cam 1 im ag e v Cam 2

Visual based model-free control

Visual model Flexible Aircraft Intelligent Controller Aeroelastic State Estimation (elastic states) input gust/turbulence visual frame [1088x600] output Real-time feedback

GeConv

Image filters and clustering (DBSCAN):

[1088x600x3] [1088x600x1] [1088x600x1] +Otsu

[1088x600x1] + Cluster

Sort and rotate points:

P_θhull = sort(P, sort(θ_cpk)); R = cos(θk) −sin(θk)

sin(θ_k) cos(θ_k)



Geometric convolution and symmetry score:

P_θk = (R · (θ_k−1 − θ_cp)T)T + θ_cp

θ_{V symm} =min(|mean(| sin(|θ_kcp|)|), θ_{V symm}) θ_Hsymm =min(|mean(| cos(|θ_kcp|)|), θ_Hsymm)

GeConv : Geometric Convolution process

Input [1088x600x3] HSV-BW [1088x600x3] BW img [1088x600x1] Contours [1088x600x1] Cluster [1088x600x1]

θ

_p 0 200 400 600 800 1000 -100 0 100 200 300 400 500 600 700 800 0 30 60 90 120 150 180 210 240 270 300 330 0 100 200 300 400 minVsmm = 0.74991 minHsmm = 0.0067772 Vsmm = 0.25002 Hsmm = 0.96794 theta = 19.5 0 200 400 600 800 1000 -100 0 100 200 300 400 500 600 700 800 0 30 60 90 120 150 180 210 240 270 300 330 0 100 200 300 400 minVsmm = 0.48985 minHsmm = 0.0067772 Vsmm = 0.51015 Hsmm = 0.85998 theta = 39.5 0 200 400 600 800 1000 -100 0 100 200 300 400 500 600 700 800 0 30 60 90 120 150 180 210 240 270 300 330 0 100 200 300 400 minVsmm = 0.22648 minHsmm = 0.0067772 Vsmm = 0.77352 Hsmm = 0.63363 theta = 59.5 0 30 60 90 120 150 180 210 240 270 300 330 0 100 200 300 400 minVsmm = 0.0067772 minHsmm = 0.0067772 Vsmm = 0.99322 Hsmm = 0.11615 theta = 97 0 200 400 600 800 1000 -100 0 100 200 300 400 500 600 700 800

DCNN structure

-θ[°] -0° -1° -2° CONV - RELU

INPUT MAXPOOL CONV - RELU MAXPOOL FLATTEN FC SOFTMAX

... ... ... ...

Extract features Classify angle

DCNN: training Deep Convolutional Neural

Network

Initial orientation Rotated Orientation Corrected Orientation

Planning Smart-X

Current work with DCNN

Measured displacement Predicted displacement Ground truth Resnet-50 architecture Update weights Error metric/ Loss function ... ... ... Smart-X