Intelligentes Inspektionssystem auf der Basis neuronaler Netze
IntroductionWithin the project Intelligent Inspection System based on Neural Algorithms new methods and algorithms as well as the software of a prototype for recognition of objects from color images have been developed at the GFaI. The system provides full self-organisation. The parametrization of the system necessary to solve a specific task is achieved by Learning from Examples. No task specific filters or feature extraction methods have to be programed a priori.Because of the algorithms used the IIS especially targets recognition tasks where objects are not or only with difficulties distinguishable by symbolic features a priori known or exact prototyp matching/measuring. Typical applications of the IIS are cases where the objects shape, coloring etc. are distributed randomly, for example the sorting of agricultural products or as intended within the project the sorting of waste packages. The realization includes two components, the training system and the real-time recognition system. Both have been implemented at an Parsytec Cognitive Computer, a high performance parallel computer. AlgorithmsA pattern recognition problem and so far the method to be used is defined by a priori knowledge on the problem.In case of object recognition from images for example if visual object properties a kown a priori distinguishing objects of different classes exactly filters can be programmed a priori to get these features. Thereafter the objects can be classified at hand of this features. That type of pattern recognition can be called exact matching or measuring (wherby the measure is a distinct a priori known distinguishing visual object property). If not all properties distingushing objects of different classes are known exactly a priori or if they can not be extracted from images exactly (f.e. noise) the task becomes to compute the probabilty that an object belongs to a class given an image of the object. The a priori knowledge is used to make assumptions on the underlying density distribution. That is done as well explicit as implicit distributed over the different parts of the recognition algorithm (taking the image, preprocessing, feature extraction, classification). If the distribution and their parameters are choosen a priori the approach is called Parametric Estimation. If only the distribution is choosen a priori and their parameters are estimated from a set of samples (pairs of images and relating classes) the approach is called Non-parametric Estimation. In that case the mapping from images to classes is estimated by Learning from Examples under some a priori assumptions. So the art of engineering is to explore properties distinguishing objects of different classes and to develop algorithms which incooperate this a priori knowledge implicit or explicit as assumption on the underlying density distribution. The applications IIS target desire a non-parametric approach. Especially for waste package sorting there can be made only a few very general assumptions. The images contain high-correlated data. That's why they can be compressed without loss of significant variance. That reduces the numerical effort of following processing steps. The compressed images called feature vectors are used as input of a classifier. It is addionally necessarry to preprocess the images to reduce noise and disturbations that is to reduce the variance of images of same classes. At the IIS images are taken from objects at a black conveyor belt iluminated by a flash light. The objects orientation at the belt is not defined. To reduce the resulting high intra-class variance a translation and rotation invariant transformation is done: First the background is removed by thresholding. Than the object is centered by shifting the mass point (center of gravity) of the relating grey image to the geometric center point. Therafter the image is rotated with respect to the first and second inertial moment of the grey image.   generic images preprocessed images As mentioned above a non-parametric approach is necessarry for the IIS. So the compression method also called feature extraction is desired to be parametrizable by Learning from Examples. The discrete Karhunen-Loéve-Transformation (KLT) is such a method under assumption of a Normal Distribution . The KLT transforms an image linearly to the orthogonal space spanned by the Eigenvectors of the estimated covariance matrix of all sample images. Because of high correlation of data in image regions it is only necessarry to use a space spanned by Eigenvectors of the n greatest Eigenvalues. An Eigenvalue represents the variance of the set of values achieved by multiplying all images of the training set with the relating Eigenvector. So the set of feature vectors achieved by multiplying the images with the Eigenvectors relating to greatest Eigenvalues will preserve 'image parts' with high variance and remove those with low variance. The underlying assumption is that objects of different classes can be distiguished by those high-variant features. Additionally to that KLT feature vector a color feature vector is computed because it is known a priori that some classes of objects are distinguishable from color. The color feature vector should give a compressed describtion of the pixel distribution in the RGB space of an image. An objects ilumination varies depending on its orientation at the conveyor belt. To reduce the resulting intra class variance only the direction of the RGB vectors (given for each pixel by the values of the red, green and blue channel) are used. The color feature vector is computed as histogram of all RGB vector directions of an image. The histogram intervals are given by prototype vectors computed by vector quantization . The vector quantization sets the prototype vectors iteratively to approximate the density distribution of the RGB vector directions of all images from the training set. Both feature vectors are used as input for the neural classifier generated by the TACOMA algorithm. RealizationWithin the IIS project a prototyp for waste package sorting has been realized. The software consists of to components, a training system and a real-time recognition system. Images are taken from separated waste packages at an conveyor belt iluminated by a flash light using a CCD camera connected to a PC running Windows NT. This host PC communicates (sends the images and receives the classification results) via a High Speed Link with the real-time recognition system implemented at an Parsytec Cognitive Computer (CC), a high performance parallel computer. The throughtput can be adapted to the application needs simply by modifying the number of nodes of the parallel computer. The algorithms of the real time recognition systems have been parallelized using Parsyframe, a Boolean Data Flow based communication layer at the CC. At the host PC the results of the visual classification are combined with sensorial information on the material provided by infrared spectroscropy and an metal detector using a decison tree. The final results are sent to the process controller unit which triggers the relating pneumatic separation actor.The training system uses images from mass storage device (hard disk, cdrom). It is equiped with an graphical user interface based on TCL/TK to control and monitor the training process. It is implemented parallel as well at a CC as at a Parsytec Power X'plorer using a SPARC station as host. A Xserver is required for both versions. ResultsThe prototyp of the IIS has been trained and evaluated with about 3500 waste package RGB images of size 376 x 280. The objects to be recognized (without considering the information of the infrared spectroscropy and an metal detector) have been:
Half of the images are used for training and the other half for independed testing of the recognition quality. Both sets are choosen to be identically distributed with respect to the generic sample set. The rate of wrong classified objects (overall error rate) is between 12.71 and 5.1% depending on the rejection threshold. The error rate varies between the classes because of the different number of objects per class (see the statistics image). The recognition matrix is well suited to evaluate the recognition quality more in detail. The next three images show this matrix for a rejection threshold of 0.0, 0.2 and 0.5. The 'error rate' row gives the percentage of objects wrong classified to the relating class. The 'rejected' collumn shows the percentage of rejected objects of each class.  
|
 |
|
| Gefördert durch den Sächsischen
Staatsminister für Wirtschaft und Arbeit
Projektpartner
|
||
Weitere Informationen |
© GFaI 1994-2005 Kontakt: Dr. H.-M. Voigt
