FAST INLINE INSPECTION OF APPLES BY A NEURAL NETWORK BASED FILTERED BACKPROJECTION METHOD. Eline Janssens

Transcription

1 iminds Vision Lab, University of Antwerp (CDE), Antwerp, Belgium Centro de Informática, Universidade Federal de Pernambuco, Recife BIOSYST-MeBios, KU Leuven, Heverlee, Belgium FAST INLINE INSPECTION OF APPLES BY A NEURAL NETWORK BASED FILTERED BACKPROJECTION METHOD Eline Janssens, Luis F. Alves Pereira, Jan de Beenhouwer, Mattias Van Dael, Pieter Verboven, Bart Nicolaï, Jan Sijbers 11/02/2016 Eline Janssens

2 Contents 2 X-ray inline inspection NN-hFBP Results Conclusions

3 X-ray Inline Inspection

4 Problem 4 Inspection of fruit: Outside Inside - New tomographic reconstruction and analysis methods - Inline inspection of food Requirements: Good reconstructions Inexpensive Fast Image from: Grant J., Mitcham B., Biasi B. and Chinchiolo S., Late harvest, high CO 2 storage increase internal browning of Fuji apples,

5 Inline Inspection 5 Inspection lines in industry Radiographic inspection source No 3D information Relative Inexpensive Fast

6 Inline Inspection 6 Tomographic inspection source Reconstruction 3D information Relative Inexpensive Slower Artefacts

7 Inline Inspection 7 Proposed inspectionmethod source Reconstruction Challenges: Fast Inspection Good Quality 3D information Relative Inexpensive Rotation

8 Classic Reconstruction 8 algorithms Fast inspection Analytical Reconstruction E.g.: FBP Very Fast Reconstruction Slow acquisition Lange number of equiangular projections Fast inspection 16 vs 128 Good quality projections Good quality Iterative Reconstruction E.g.: SIRT Slow reconstruction Fast acquisition Small number of projections Incorporate Prior Knowledge 16 vs 128 projections

9 Inline scanning geometry 9 Flatpanel Detector source Virtual Moving Detector source

10 Neural Network Hilbert transform based Filtered Back Projection

11 NN-FBP : Neural Network FBP Parallel beam Circular scanning geometry Contribution Fan beam: hfbp Inline scanning geometry Pelt D.M., Sijbers J., Batenburg K.J.,2013a, Fast Tomographic Reconstruction from Highly Limited Data Using Artificial Neural Networks, Proceedings of 1 st International Conference on Tomography of Materials and Structures (ICTMS 1), Ghent, Belgium,

12 NN-FBP 12 Output of a Neural Network: Network trained to reconstruct 1 pixel z 1 n,,, σ σ b b OUTPUT q 0 w 0 q 1 FILTERINPUT WEIGHT z 2 z 3 w 1 w 2 w 3 q 2 q 3 z N σ = activation function b = bias Filters and weights are trained by the neural network

13 NN-FBP 13 Output of a Neural Network: n,,, σ σ b b FBP x,y z 1 hτ & P '( x cos θ./ sinθτ & z 2 w 0 ' ( w 1 q 0 q 1 z 3 z N w 2 w 3 q 2 q 3

14 NN-FBP 14 Output of a Neural Network: n,,, σ σ b b FBP w0 q 0 FBP w1 q 1 q 2 FBP w2 q 3 FBP w3

15 Fan-beam 15 Option 1: Convert fan beam to parallel beam FBP w0 q 0 Fan Beam FBP w1 q 1 Parallel Beam FBP w2 q 2 q 3 FBP w3

16 You J., Zeng GL., Hilbert transform based FBP algorithm for fan-beam CT full and partial scans, IEEE Trans Med Imaging, 26(2), 190-9, February 2007 Fan-beam 16 Option 2: Work directly on fan-beam data Hilbert transform based FBP hfbp w0 q 0 Fast Good reconstructions Directly on Fan-beam da Filter x Input hfbp w1 q 1 q 2 hfbp w2 q 3 hfbp w3

17 Results

18 Specifications 18 Data Specifications Jonagored apples with defects Reconstruction 256 x 256 Simulate inline projections from available circular projections Moving detector 287 pixels Data from BIOSYST-MeBios, KU Leuven, Belgium Network Specifications Training: 100 images from 9 apples Validation: 10 images from 9 different apples Test: 50 images from 5 apples random pixels for training and for validation 4 Hidden Nodes

19 19 Non-Equiangular Projections

20 Reconstructions projections projections projections

21 Error images projections FBP SIRT NN-hFBP 64 projections FBP SIRT NN-hFBP

22 22 Comparison Projections NNhFBP NNhFBP RMSE RMSE mask MAD FSIM #proj EA N-EA EA N-EA EA N-EA EA N-EA First Indication: Similar quality achieved for both equiangular and non-equiangular projections

23 23 Reconstruction time

24 Conclusions

25 Conclusions 25 Conveyor belt tomography with a fixed source-detector system and an additional object rotation is feasible. NN-hFBP outperforms both FBP and SIRT in terms of image quality. The reconstruction time of NN-hFBP is an order of magnitude smaller compared to SIRT.

26 26 Thank you for your attention Agency for Innovation by Science and Technology in Flanders (IWT SBO TomFood)