Comparison of Lamstar NN & Convolutional NN – Character Recognition

Comparison of Lamstar NN & Convolutional NN – Character Recognition

D Gaupe, C Contaldi, A Sattiraju,
University of Illinois, Chicago. IL, USA

SUMMARY OF RESULTS
_________________________________________________________________
CNN CNN LNN-1 LNN-2
_________________________________________________________________

8x8 10x10 8x8 8x8

Training
/iteration 159 ms 3.6 ms 3.6 ms

Testing/it 34,74 ms 0.04857 ms 0.04857ms

Performance %
in noise
0 bits 100 100 100 100
1 bit 100 100 100 100
2 bits 100 100 100 100
3 bits 100 100 100 100
8 bits 100 100 100 100
10 bits 100 100 100 100
13 bits 100 100 100 100
15 bits 100 98 98.44 100
16 bits 98 92 100 100
20 bits 94 94 100 100
25 bits 92 82 89.06 95.31
30 bits 78 70 78.13 87.5
40 bits 72 58 59.56 79.56
45 bits - - 29.69 60.94
48 bits - - 28.13 48.44

Language MATLAB MATLAB MATLAB MATLAB

COMMENT:
The CNN results are based on a CNN algorithm based on [2]. LAMSTAR -1 (LNN-1) is described in Chapter 9 of [1]. LAMSTAR-2 (LNN-2 above, also known as Modified LAMSTAR) is described in Section 9.8 of [1]. The LAMSTAR-1 and LAMSTAR-2 programs use 16 input layers, 8 neurons per layer. Each layer represents a scan of one row or one column of the input image.







The above results are taken from Homework Project, ECE/CE 559, UIC (Professor: D. Graupe), Oct. 2015.

The results above are from a project was carried out by Carlo Contaldi , Grad Student, ECE Dept., UIC (class ECE/CS 559).


REFERENCES

[1] LAMSTAR & Modified LAMSTAR (LAMSTAR 2): see - Chapter 9 in: D Graupe, Principles of Artificial Neural Networks, World Scientific Publishers, 3rd Edition, 2013

[2] Mihail Sirotenko "CNN Convolutional Neural Network Class" 2009:
http://www.mathworks.com/mathlabcentral/fileexchange/24291-cnn-convolutional-neural-network-class