Application of neural networks into automatic visual diagnostic of railway wooden sleepers

Abstract: The paper presents the system composing of neural network and image processing procedures being able to classify wooden sleepers on the basis of their image. Image processing procedures extract salient features of sleeper that are further used by neural network in classification process. The system performance was checked on 100 images of good sleeper and 100 images of bad sleeper. System classification rate was equal to 84% for images not taking part in learning process, and 95% for images taking part in learning process.
1. INTRODUCTION
Good condition of wooden sleepers is a crucial problem having a large influence on the safety of railway transport. Every year, poor condition of wooden sleepers poses potentially threat to railway traffic. They can cause train derailment, what in turn can generate tremendous human and financial losses. At present, in Polish Railway Lines skilled persons manually check wooden sleeper’s condition. They perform visual inspection of wooden sleepers. Such approach is highly inefficient and dependent of the fatigue of controlling person, one person can check daily very few sleepers. Additionally, do not exist any regulations precisely describing the difference between the good and bad sleeper. The inspection is based on the experience and knowledge of controlling person 14.
These disadvantages make authors deal with elaboration of automatic system serving to inspection of wooden sleepers. Needless to say that, it is not a simple task. This system should be immune against pebbles occurring on sleepers and variable texture of sleepers.
In order to satisfy such tough requirements, it is necessary to use the fusion of image processing procedures and neural networks. Authors intention is not the creation of complete automatic system, but rather focusing on the choice of the best inspection algorithm and checking its usefulness to automatic inspection of wooden sleepers. This system will consist of hardware and software. In our experiment the hardware is confined
to the camera. Pictures of both good and bad sleepers have been taken with this camera.
The overall number of pictures being taken is equal to 200, 100 corresponding to good sleepers and 100 corresponding to bad sleepers. The software consists of image processing procedures and classification algorithm (neural network). Image processing procedures also called preprocessing serve to extract the salient features of the sleeper best separating bad sleepers from good sleepers. (...)
Więcej w tej kategorii: « Henry Lloyd - łańcuchy zaopatrzenia Technologia realizacji usługi komunikatów dla kierowników statków (NtS) w systemach RIS »
Zaloguj się by skomentować