L. Spaanenburg

Car license plate recognition with neural networks and fuzzy logic

A car license plate recognition system (CLPR-system) has been developed to identify vehicles by the contents of their license plate for speed-limit enforcement. This type of application puts high demands on the reliability of the CLPR-system. A combination of neural and fuzzy techniques is used to guarantee a very low error rate at an acceptable recognition rate. First experiments along highways in the Netherlands show that the system has an error rate, of 0.02% at a recognition rate of 98.51%. These results are also compared with other published CLPR-systems.

License plate recognition using DTCNNs

Automatic license plate recognition requires a series of complex image processing steps. For practical use, the amount of data to be processed must be minimized at early stage. This paper shows that the computationally most intensive steps can be realized by discrete time cellular neural networks (DTCNNs). Moreover, high-level operations like finding the license plate in the image and finding the characters on the plate need only a small number of DTCNNs. Real-life tests show that the DTCNNs are capable of correctly identifying more than 85% out of all license plates while leaving only 0.5% of the original information to be inspected for actual recognition.

Transformational DT-CNN design from morphological specifications

Morphology provides the algebraic means to specify operations on images. Discrete-time cellular neural networks (DT-CNNs) mechanize the execution of operations on images. The paper first shows the equivalence between morphological functions and DT-CNNs. Then, the argument is extended to the synthesis of optimal DT-CNN structures from complex morphological expressions. It is shown that morphological specifications may be freely derived, to be subsequently transformed and adopted to the needs of a specific target terminology. This process of technology mapping can be automated along the well-trodden path in CAD for microelectronics.

On the representation of data for optimal learning

The key to the successful training of a neural network lies in the careful composition of the learning set. A simple method of data screening is described to verify conformance with the restrictions imposed by the target network topology and ensuing applied to the nondestructive ultrasonic diagnosis of spot welds. It is illustrated how such a screening may indicate the suitability of a learning set before application to the neural net and therefore aids in designing a proper precoder.

Assembling engineering knowledge in a modular multi-layer perceptron neural network

The popular multilayer perceptron (MLP) topology with an error-backpropagation learning rule doesnt allow the developer to use the (explicit) engineering knowledge as available in real-life problems. Design procedures described in literature start either with a random initialization or with a smart initialization of the weight values based on statistical properties of the training data. This article presents a design methodology that enables the insertion of pre-trained parts in a MLP network topology and illustrates the advantages of such a modular approach. Furthermore we will discuss the differences between the modular approach and a hybrid approach, where explicit knowledge is captured by mathematical models. In a hybrid design a mathematical model is embedded in the modular neural network as an optimization of one of the pre-trained subnetworks or because the designer wants to obtain a certain degree of transparency of captured knowledge in the modular design.

Learning feed-forward multi-nets

Multi-nets promise an improved performance over monolithic neural networks by virtue of their distributed implementation. This potential lacks popularity as, without precautions, the learning rate has to drop considerably to eliminate the occurrence of unlearning. This paper introduces extensions of the Error Back-Propagation algorithm to enable function preserving merging of neural modules at full learning rate.

Over multiple rule-blocks to modular nets

Real production data are vague and irreproducible. This suggests fuzzy knowledge acquisition for an online learned combined fuzzy/neural network. The paper advocates a modular neural only network based on the injection of knowledge from a multiple rule block fuzzy specification. A typical controller takes 10-25 blocks with 10 rules on average, which after injection in a neural net can be personalized and adapted by small (typical 100) measurement sets.

Design of discrete-time cellular neural networks based on mathematical morphology

Mathematical morphology is a discipline that provides a formal framework for the analysis and manipulation of images. Its theoretical foundations have been well-established in the last forty years and it has shown to be a power fool tool in the development of a large number of image processing applications. This paper shows that a lot of knowledge that is developed in the field of mathematical morphology can be applied to discrete-time cellular neural networks (DTCNNs). DTCNN equivalencies of the elementary morphological operators, which are the basic building blocks for complex image operations, are introduced and the correctness of these templates is formally proved.

CNN-Applications in Toll Driving

This paper describes a system for the automatic identification of vehicles by the contents of its license plate. The system has been developed over the last five years and is one of the four candidate systems for automatic toll collection in the Netherlands. Due to the extreme time and performance requirements placed on the system, advanced techniques like DT–CNNs and classifier combining are used. The first tests on the road show that the system correctly recognizes 85.4% of the passing vehicles, while marking the remaining 14.6% as unrecognizable.

Efficient DTCNN implementations for large-neighborhood functions

Most image processing tasks, like pattern matching, are defined in terms of large-neighborhood discrete time cellular neural network (DTCNN) templates, while most hardware implementations support only direct-neighborhood ones (3/spl times/3). Literature on DTCNN template decomposition shows that such large-neighborhood functions can be implemented as a sequence of successive direct-neighborhood templates. However, for this procedure the number of templates in the decomposition is exponential in the size of the original template. This paper shows how template decomposition is induced by the decomposition of structuring elements in the morphological design process. It is proved that an upper bound for the number of templates found in this way is quadratic in the size of the original template. For many cases more efficient and even optimal decompositions can be obtained.