AIMS Management Handbook
Table of contents
1.
Task description.
2.
Executive Summary.
3.
Video Library for Feature Extraction.
3.1.
Introduction.
3.2.
Visual Library and initial conditions.
3.2.1.
Video Library images requirements.
3.2.2.Video Library Optimization.
3.2.3.Size optimization process.
3.2.4.Labeling (Naming)
and storing the pictures in database.
3.2.5.Video Library Automation data support system..
3.2.6.User Interface and functionality.
4.
Mathematical Approaches for features extraction.
4.1
Introduction and basic requirements.
4.2
Image processing possibility for mathematical approaches.
5. Feature Extraction and
Classification Methods.
5.1.1. Color distribution methods.
5.1.1.1. RGB Method.
5.1.2. Texture analyzing methods.
5.1.2.1. Histograms.
5.1.2.2.Moments.
5.1.2.3.Co-occurrence matrix 32*32 pixels.
5.1.2.4.Co-occurrence matrix 64*64 pixels.
5.1.2.5.Markov Random Fields- Energy Distribution.
5.1.2.6.Markov Random Fields Gibbs middle
intensity-Energy Distribution.
5.1.2.7.Threshold.
5.1.2.8.MRF-Threshold.
5.2.Classification of information extracted from the methods.
5.3.Weights according to the probability and the time of processing.
5.4.CONCLUSIONS..
6. Neural Network.
6.1.Theoretical description of neural nets method calculation.
6.1.1.Introduction.
6.1.2.Cellular neural network.
6.1.3.Self- organizing mechanism system..
6.1.4.Basic Profits of using Cellular Neural Networks.
6.1.5.Profits from a self organizing mechanism..
6.1.6.General Description of
used Cellular Neuronal Network
6.1.7.Brief description of the layers.
6.1.7.1.Layer 1.
6.1.7.2. Layer 2.
6.1.7.3.Layer 3.
6.1.7.4.Layer 4.
6.1.8.Decision Cells.
6.1.9. Database of Knowledge.
7. The block scheme of the system and user interface.
8. Conclusions.
8.1.Mathematics models and data processing.
8.2.Software.
8.3.Industrial realization.
CONCLUSIONS
We can make some
conclusions in order to complete this final report
This task represents a
complicated research work cross-section between mathematics, software, and
engineering.
Using this can be
extracted several classes conclusions according to the task requirements.
·
The result form corrosion zone
separating shows that between 0 and 1500 hours the recognition is more
precise than in 1500-8000 hours zones.
·
The mathematical models should
be combined to be able to produce data in 0-8000 hours corrosion range.
·
The calculation time is not
proportional always to the probability of the result.
·
The Cellular Network is the
optimum approach combining self-organizing system and reducing total
calculation time.
·
The model is very sensitive to
initial learning process and initial data library
·
Exclude the interface
(prepared for scientific purposes) the basic functions and principles are
ready for industrial application.
·
The realized program proves
the idea for probabilistic corrosion categorizing using the chosen methods
and self-organizing mechanism.
·
The recognition speed is
reliable for industrial purposes.
·
The system is independent
according to light source intensity
·
For industrial application the
software should be rewritten in full C++ program, combining learning process
from video library and recognition process from Frame grabber and real
camera.
·
User interface should be
prepared depending how the system will be used (robotic or manual
realization).
|
TASK 2: Non-destructive
Testing strategies for CLASSIFICATION OF EXISTING COATED and prepared steel
SURFACES
Objectives:
Development of non-destructive testing (NDT) methods which will enable
quantified classification of the condition of existing coated and prepared
steel surfaces as an intrinsic component of an automated man-machine system.
For the existing surfaces, these methods will include classification of
substrate corrosion and coating defects such as disbonding and delamination.
Three main
design criteria are crucial for these systems:
Large
surface areas must be covered quickly.
The sensor
systems must be suitable for use in an automated man-machine system.
Currently
recognised international standards for surface condition are to be used as
references.
Rationale:
Currently, photographic standards are used for classification of the level
of corrosion of coated steel surfaces. Similarly, photographic standards are
available for levels of cleanliness achieved by the established surface
preparation methods i.e. hand cleaning with tools and dry, slurry and
ultra-high pressure water blasting. These methods of testing have the
following disadvantages [24]:
·
Both sets
of visual standards are difficult to use due to differing appearance of
steel, hue and lighting effects.
·
Destructive
testing methods must also be used to support coating condition assessment.
·
Only spot
inspection is practical.
This task
overcomes these disadvantages by developing a non-destructive, non-contact
automated surface classification system that improves classification
reliability and enables large surface areas to be quickly tested.
Task
Description:
The main classification methods will be grey scale and colour image
processing techniques [6,8] and infrared thermography synchronised with a
controllable heat source [7,19,20,21,22,23]. Artificial intelligence
techniques will be applied to the interpretation and quantifying
classification problems. A comprehensive set of manually classified (to
international standards) reference samples (80 - 100) will be made available
for this work, representing levels of substrate corrosion and corresponding
levels of surface cleanliness with different preparation methods. The
approaches in the investigation will be:
(i)
Application of colour segmentation to the detection of the pigmentation
anomalies which are frequently symptoms of substrate corrosion and coating
failure. Quantification is expected to be mainly on the basis of affected
area and location.
(ii) The
textural properties of the range of samples will be studied. Texture
analysis, including statistics and signal processing and filtering methods,
will be applied and the most reliable sets of discriminating features
identified for each class of surface. Features such as uniformity of energy
entropy, contrast, homogeneity and roughness will be extracted from
cooccurrence, grey level dependence, grey level gap length and grey level
run length matrices of the textured images. Features thus obtained will be
tested for reduction using techniques such as principle component analysis
and neural networks. The optimum basis for classification will be
investigated using classification schemes such as regression trees, neural
networks, genetic algorithms and bayesian classification. Controlled
lighting and techniques for invariance with illumination will be tested.
(iii)
Lockin thermography with modulated lamps will be used to inspect samples
representing delaminations, paint disbondings and corrosion. Measurements of
paint thickness and corrosion thickness will also be performed with lockin
thermography on the base of reference samples. Depending on the kind of
problem, validation techniques such as ultrasonics and eddy currents will be
used in order to assess the quality of results obtained using the lockin
thermography technique. The partner responsible for providing the samples
(see Task 1) will know in advance the level of corrosion of every sample.
This "a priori" knowledge will allow the validation of lockin thermography
results for determination of corrosion levels, thus avoiding the need of
highly expensive destructive tests. Porosity evaluation even if possible,
will not be performed due to the high costs related to the validation
procedure. |
