Modified quadratic discriminant functions and the application to chinese character recognition

Area of Science:

• Pattern Recognition
• Machine Learning
• Statistical Classification

Background:

• Quadratic Discriminant Functions (QDF) are widely used for classification.
• Estimation errors in covariance matrices can degrade QDF performance.
• Robustness to estimation errors is crucial for reliable discriminant analysis.

Purpose of the Study:

• To address the sensitivity of Quadratic Discriminant Functions (QDF) to covariance matrix estimation errors.
• To propose novel Modified Quadratic Discriminant Functions (MQDF1 and MQDF2) with enhanced robustness.
• To evaluate the effectiveness of MQDF1 and MQDF2 in a practical application like Chinese character recognition.

Main Methods:

• Proposed MQDF1 utilizes a pseudo-Bayesian estimate for the covariance matrix, differing from the standard maximum likelihood estimate.
• Developed MQDF2 as a computationally efficient variation of MQDF1, reducing time and storage requirements.
• Applied both MQDF1 and MQDF2 to a Chinese character recognition task for performance evaluation.

Main Results:

• MQDF1 and MQDF2 demonstrated reduced sensitivity to covariance matrix estimation errors compared to standard QDF.
• Significant performance improvements were observed when applying MQDF1 and MQDF2 to Chinese character recognition.
• MQDF2 offered a favorable trade-off between accuracy and computational efficiency.

Conclusions:

• The proposed Modified Quadratic Discriminant Functions (MQDF1, MQDF2) offer a more robust alternative to traditional QDF.
• These methods effectively mitigate the impact of covariance matrix estimation errors in classification tasks.
• MQDFs show considerable promise for improving the accuracy and efficiency of pattern recognition systems, particularly in character recognition.