Hybrid Quantum–Classical versus Classical Convolutional Neural Networks for Sign-Like Digit Recognition

This paper compares a classical convolutional neural network (CNN) with a hybrid quantum–classical convolutional neural network (QCNN) that appends a simulated parameterized quantum circuit to a convolutional front end. Both models were implemented with TensorFlow, TensorFlow Quantum and PennyLane and trained for ten epochs (batch size 64) on MNIST, used as a simplified proxy for numerals on speed limit signs. The hybrid model reached 98.67% accuracy after ten epochs versus 96.64% for the CNN, rose faster in early epochs, and had a median time of 139 s per epoch versus 440.5 s. Because the models differ in classical depth and were each trained once, the results do not establish a quantum advantage. The paper also discusses deployment barriers in vehicles and gives a SWOT comparison.