55.

[Cover] [Contents] [Index]

Page 147

Table 3.3 Confusion matrix output by multilayer perceptron, SOM, counter-propagation and fuzzy ARTMAP network

No.	1	2	3	4	5	6	7	u.a%.
(a) Multilayer perceptron
1	9580	5171	852	578	2205	871	61	49.59
2	790	16644	2229	439	1021	569	26	76.64
3	236	7302	5554	283	914	541	22	37.40
4	147	607	93	10587	5285	305	7	62.16
5	525	1149	180	2489	17051	585	66	77.35
6	297	2125	695	599	2137	7308	16	55.46
7	293	1040	171	403	1529	162	1578	30.49
p.a.%	80.72	48.90	56.82	68.85	56.57	70.67	88.85	68302
(b) SOM
1	9340	4255	621	648	2664	720	51	51.04
2	752	16750	1984	376	1464	597	31	76.30
3	468	7916	5947	483	1252	599	11	35.66
4	313	1129	216	11114	4298	344	10	63.79
5	435	1065	204	1956	17397	767	48	79.54
6	378	2253	731	514	1762	7165	20	55.88
7	182	670	71	287	1305	149	1605	37.60
p.a.%	78.70	49.21	60.85	72.27	57.72	69.29	90.37	69318
(c) Counter-propagation
1	8139	4799	741	981	4568	796	55	40.53
2	810	12084	1971	408	872	441	25	72.75
3	345	9111	5138	481	1429	713	45	29.76
4	444	1695	249	9626	5645	646	12	52.55
5	1206	1456	354	2124	12611	862	74	67.49
6	557	2995	915	1517	3951	6621	64	39.84
7	367	1898	406	241	1066	262	1501	26.15
p.a.%	68.58	35.50	52.57	62.60	41.84	64.03	84.52	55720
(d) Fuzzy ARTMAP network
1	8056	3120	392	316	1737	360	36	57.48
2	824	14837	1678	425	1506	525	16	74.89
3	591	8584	5718	405	898	535	10	34.16
4	442	1139	265	10480	4227	247	13	62.33
5	706	1593	336	2513	16420	638	35	73.83
6	892	3737	1149	904	3475	7809	39	43.37
7	357	1028	237	335	1879	227	1627	28.59
p.a.%	67.88	43.59	58.50	68.15	54.48	75.51	91.61	64947
(a) Total Accuracy: 60.28%; kappa coefficient: 0.521. (b) Total Accuracy: 61.17%; kappa coefficient: 0.531. (c) Total Accuracy: 49.17%; kappa coefficient: 0.397. (d) Total Accuracy: 57.31%; kappa coefficient: 0.488.

The results of these experiments indicate that the multilayer perceptron is a suitable tool for image classification in terms of both classification accuracy and training time. Although the performance of the SOM networks is comparable to that of the multilayer perceptron, its considerably

[Cover] [Contents] [Index]