Quantum Classification Methods for Optical Character Recognition: Unlocking Improved OCR Performance in Limited Data Regimes

This paper examines the applicability of quantum machine learning methods to optical character recognition (OCR). The paper's abstract is given below: This paper builds on previous research demonstrating the effectiveness of quantum machine learning methods on limited or restricted data sets. A simulated quantum circuit is used to determine the validity of quantum machine learning methods for optical character recognition (OCR). OCR models are typically implemented using Convolutional Neural Networks (CNN), which can require large data sets to be trained effectively. A demonstrated ability to produce accurate OCR results with quantum machine learning would enable OCR-like models to be used when only limited data is available. This paper compares the effectiveness of quantum machine learning and classical CNN models on restricted samples of the MNIST data set, implemented in Python using IBM's Qiskit library and TensorFlow. Due to better generalization and feature space expressivity, the quantum model outperformed the CNN by an average of 6.33% on training data sizes of 10 samples to 300 samples. Both predictions had equal accuracy at $N=300$ samples, suggesting the CNN will either overtake or remain equal to the accuracy rate of the QSVC for larger data sample sizes. The space used by each model was explicitly measured, with the QSVC utilizing approximately 86% less memory space than the CNN. Upon evaluation of kernel heatmaps and eigenvalue spectra, performance improvements were minimal past $N>200$ samples for training data. The time complexity of this QSVC model has a minimum of $\Omega(N^2)$, so the time cost outweighs any performance benefit. The ability to run OCR models with reliable accuracy for small data sets creates possibilities for on-device OCR models, launching new opportunities for highly specialized edge models. This will evaluated in future research.