E.K. Teoh

2005 Special Issue: Generalized 2D principal component analysis for face image representation and recognition

In the tasks of image representation, recognition and retrieval, a 2D image is usually transformed into a 1D long vector and modelled as a point in a high-dimensional vector space. This vector-space model brings up much convenience and many advantages. However, it also leads to some problems such as the Curse of Dimensionality dilemma and Small Sample Size problem, and thus produces us a series of challenges, for example, how to deal with the problem of numerical instability in image recognition, how to improve the accuracy and meantime to lower down the computational complexity and storage requirement in image retrieval, and how to enhance the image quality and meanwhile to reduce the transmission time in image transmission, etc. In this paper, these problems are solved, to some extent, by the proposed Generalized 2D Principal Component Analysis (G2DPCA). G2DPCA overcomes the limitations of the recently proposed 2DPCA (Yang et al., 2004) from the following aspects: (1) the essence of 2DPCA is clarified and the theoretical proof why 2DPCA is better than Principal Component Analysis (PCA) is given; (2) 2DPCA often needs much more coefficients than PCA in representing an image. In this work, a Bilateral-projection-based 2DPCA (B2DPCA) is proposed to remedy this drawback; (3) a Kernel-based 2DPCA (K2DPCA) scheme is developed and the relationship between K2DPCA and KPCA (Scholkopf et al., 1998) is explored. Experimental results in face image representation and recognition show the excellent performance of G2DPCA.

Programming Hopfield network for object recognition

This paper investigates the performance of the Hopfield neural network as a constraint satisfaction network for invariant pattern recognition. Although the Hopfield network is known to provide instantaneous solution to optimization problems with combinatorial complexity, in some instances the solution is invalid. In this paper, we study a number of energy function formulations and experimentally explore their merits. We also present an industrial application of Hopfield network in recognizing transparent flexible membrane printed circuits and a subgraph isomorphism of synthetic line patterns invariant of position, scale and orientation. The proposed network can correctly recognize overlapped partial line patterns and offers highly parallel implementation.<<ETX>>

An intelligent robotic vision system for inspection of surface mount PCBs

The authors present techniques used for defect inspection on surface-mount printed circuit boards (PCBs). The focus is on five different types of defects, namely, missing components, misalignment, wrong orientation of IC chips, wrong parts, and poor solder joints. Five separate algorithms have been developed to detect these faults. The technique of windowing was used to reduce the amount of redundant data to be processed. Preprocessing functions like convolution as well as all image processing tasks are performed on the window regions only, saving a tremendous amount of computation time. Experimental results showed that these algorithms are reliable, fast and cost-effective.<<ETX>>

Fuzzy connectives based optimal mapping of homomorphic ARG matching onto self-organising Hopfield network

Attributes used in object recognition can be considered fuzzy variables as they are generally noisy, unreliable and ambiguous. In this paper, the authors employ fuzzy information aggregation operators to optimally map the attributed relational graph (ARG) matching problem onto the self-organising Hopfield network. The computation of the parameters used in the information aggregation operators is formulated as a constraint optimization problem and solved using the gradient projection based learning algorithm. The mapping scheme ensures that the problem is optimally mapped for every model. Experimental results clearly show the usefulness and necessity of the learning scheme.

A method for semi-quantitative model-based reasoning system

Describes a method for creating a semi-quantitative model-based reasoning system for ECG (electrocardiogram) signal interpretation. By this method, a quantitative cardiac model based on the principle of cardiac vectorial analysis was built into the interpretation system. The system makes a diagnostic interpretation on the input ECG signals by reasoning about the model through model simulation. The system may also combine both associational-level knowledge and deep-level causal knowledge from the semi-quantitative model, so as to overcome some limitations of the model-based reasoning system.<<ETX>>

On mapping of ARG matching onto neural networks

Learning schemes are presented to optimally map the homomorphic graph matching problem onto the Potts mean field theory neural networks. The computation of the weighting factors used in the compatibility measure equation is formulated as an optimization problem and solved using the quadratic programming procedure based learning algorithm. The formulation implicitly evaluates ambiguity, robustness and discriminatory power of the relational attributes chosen for graph matching and assigns weighting factors appropriately to these relational attributes. Further, the tolerance and steepness parameters are also learnt. These learning schemes also enable us to construct the augmented weighted model attributed relational graphs (WARG). The proposed parameter learning schemes are employed to solve the silhouette objects recognition problem and the necessity for such learning schemes is demonstrated.

Learning critical temperature for homomorphic ARG matching by self-organising Hopfield network

The authors previously (1995) presented a programming strategy to generate a homomorphic mapping between two attributed relational graphs (ARG) by the Hopfield network. Further, a self-organisation scheme was also introduced to learn the constraint parameter used in the energy function. In order to generate the desired mapping, the temperature parameter should be set to the critical value. Estimation of the critical temperature is an extremely difficult problem. In this paper, a heuristic learning algorithm is presented to estimate a suitable value for the temperature parameter for every model and scene pair to be matched. Experimental results showed that the learning algorithm is capable of compensating for the variations in the model and scene characteristics and the time step used to simulate the dynamic equations of the Hopfield network.

Intelligent control of robotic manipulators with DSP

Recent advances in computing technology have been providing the fundamental driving force for the sudden increase in interest in intelligent control. In this paper an autonomous self-organizing fuzzy logic controller is formulated and implemented using a DSP-based controller card to control in multiplex three links, namely base, shoulder and elbow, of a revolute joint manipulator. Simulation and online testing results have verified the self-optimizing ability of the process independent controller. The controller presented here can be easily formulated and implemented using low-cost DSP chips.<<ETX>>

Programming Hopfield network for relational homomorphism

We study the energy and compatibility function formulations for pattern recognition by homomorphic mapping of attributed relational graphs using the Hopfield network. A deterministic hypothesis initialization strategy is introduced and proven to be superior to the commonly used random initialization in many aspects. Further, a method to verify the validity of the hypotheses generated by the Hopfield network is also presented based on a compatible cluster formation procedure using binary compatibility measures. The compatible cluster formation method allows multiple hypotheses to be evaluated simultaneously and the best to be chosen. The performance of the homomorphic algorithm is evaluated using silhouette images.<<ETX>>

A DSP-based adaptive controller for robotic control application

The development and implementation of digital controllers for position control of a DC motor using the digital signal processor (DSP) chip TMS320E15 has provided encouraging results. Real-time position control of the DC motor was investigated by using the controllers with a shaft encoder. Experimental results show that the system responses are good and are as expected theoretically. Adaptive control has become a viable alternative for controlling electromechanical systems. The adaptive control system provides the desired performance throughout the life of the mechanism. Furthermore, reliability is enhanced since the system meets the expected performance in spite of aging.<<ETX>>