Conic section function neural network circuitry for offline signature recognition

Area of Science:

• Artificial Intelligence
• Computer Engineering
• Signal Processing

Background:

• Multilayer perceptron (MLP) and radial basis function (RBF) networks offer distinct advantages.
• A unified framework can leverage the strengths of both MLP and RBF networks.
• Offline signature recognition requires robust and efficient pattern recognition systems.

Purpose of the Study:

• To design and implement a conic section function neural network (CSFNN) circuitry for offline signature recognition.
• To create a unified neural network framework combining MLP and RBF network benefits.
• To develop a general-purpose neural network circuit applicable to various pattern recognition tasks.

Main Methods:

• A mixed-mode circuit implementation was used to develop the CSFNN circuitry architecture.
• The CSFNN circuit system was designed to be problem-independent, supporting network sizes up to 16-16-8.
• A chip-in-the-loop learning technique was employed for training to mitigate analog process variations.

Main Results:

• The CSFNN circuitry was successfully applied to two distinct signature recognition problems.
• The hardware implementation achieved computational performance comparable to CSFNN software.
• The system demonstrated effectiveness in nonlinear signature recognition tasks.

Conclusions:

• The developed CSFNN circuitry provides a viable hardware solution for offline signature recognition.
• The unified framework and mixed-mode implementation offer flexibility and efficiency.
• Chip-in-the-loop learning effectively addresses analog hardware variability, ensuring reliable performance.