Considering defects of hot rolled strips are hard to be recognized by human eyes under high speed and strong noise disturbance circumstances, automatic recognition technique is discussed, and traditional space-domain feature set is very hard to be extracted, spectrum images of defects are researched in depth and are firstly extracted three types of features including Crisscross Region Feature Set which contains 240 origin features, Sum of Valid Pixels (SVP) and Repletion Ratio of Centre Region (RRCR), secondly optimized the feature set by genetic algorithm to get a bran-new feature set including 54 features. Based on the bran-new feature set, several recognition experiments of neural network are implemented, and the results show that the recognition effects under this new feature set can meet the demands of surface defect recognition of hot rolled strips in real- world under hard circumstances. Meanwhile research orientation of surface defect recognition of hot rolled strips is pointed out in order to improve recognition effects.
[1]
Randy L. Haupt,et al.
Practical Genetic Algorithms
,
1998
.
[2]
K. S. Narendra,et al.
Neural networks for control theory and practice
,
1996,
Proc. IEEE.
[3]
Dirk Weuster-Botz,et al.
Genetic algorithm for multi-objective experimental optimization
,
2006,
Bioprocess and biosystems engineering.
[4]
Chang Wook Ahn,et al.
On the practical genetic algorithms
,
2005,
GECCO '05.