Loading...
Please wait, while we are loading the content...
Similar Documents
Classiier Recognition Rate. Misclassiication. Ambiguous Rejection Rate. Distance Rejection Rate 4. Application 4.1. Data Sets Classiier Recognition Misclassiication Rejection 5. Conclusion 3.2. Neural Network Fusion Methodology Multi-classiiers Neural Network Fusion versus Dempster-shafer's Orthogon
| Content Provider | Semantic Scholar |
|---|---|
| Author | Bollivier, De Gallinari, Patrick Thiria, Sylvie Hull, Jonathan J. Michie, Donald Spiegelhalter, David J. Taylor, C. C. X=y Z. Loonis, Pierre Zahzah, El-Hadi Bonnefoy, Jean-Pierre |
| Copyright Year | 1995 |
| Abstract | Table 1: results from scene analysis, with fusion. Contribution a la repr esentation des connaissances et a leur utilisation pour l'interpr etation automatique des images satellites. The system can be seen as multi-decision point of view for the following reasons : diierent levels : baye's rule is a measurement level classiier while MLP is a rank level classiier, diierent approaches : bayesian approach is statistically driven while MLP is a NN approach, diierent inputs : each classiier is linked with a speciic parameter sub-space. The system operates in three main steps. 1. Each SPU is separately learned, it eventually transforms its outputs to conndence measurement wether decisions are produced. At the end of this stage, all SPU must be operational. 2. The NN fusion module is learned using as input the outputs given by the previous operational SPU which themselves use the same learning set as (1). 3. The system is wholly operational, it can be used for generalization. Actually, each module (SPU) is trained sepa-ratly, with a given training set. The NN fusion module is trained on the results of the SPUs. The clustering due to the MLP creates a kind of meta-frontiers of decision. It learns the coherence of the SPUs on given classes and decides which SPU is right when diierent sources disagree. For comparison, the DS's fusion treats information (outputs of each SPU) with more semantic concepts. Each decision is mapped into a conn-dence measurement and the global decision is made among a coherence notion. Because no learning allows a speciication of fusion module, the nal decision is taken on a relative coherence aspect among the diierent sources, at any time. It is well known that the problem of real application is mainly due to complexity of the information to be treated. In PR problem, various brightness, non modelizable objects, natural evolution are factors that involve perturbations of the descriptors. A contrario, this complexity involves important useful information ux which is used in our system. We have tested our system on diierent data sets issued from 4]. This data, references of whose can be found in 4], describes 7 exterior images it contains 7 classes (brickface, sky, foliage, cement, window, path, grass) and 19 parameters describing each class. According to Principle Component Analysis we have kept 12 continuous variables, split into 3 sub-spaces. A learning set is composed with 70 samples per class , they have … |
| File Format | PDF HTM / HTML |
| Alternate Webpage(s) | http://www-l3i.univ-lr.fr/~ploonis/paper/icnn95.ps.gz |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
