Surface-Crack Detection with Quantum Machine Learning

Problem Statement

Detecting surface cracks is crucial for ensuring the quality of railway tracks and the safety and durability of civil structures like bridges and buildings. Traditional method involves manual inspections or high-compute classical models are limited by scalability, hardware dependency, and accuracy under real-world conditions. Tata Elxsi explored the use of Quantum Machine Learning (QML) and Quantum Convolutional Neural Networks (QCNN) to overcome these constraints, enabling contactless, real-time, and edge-deployable models for visual inspection. 

Challenge

The project posed three main challenges,

• First, encoding high-dimensional image data into a quantum-friendly format without compromising texture fidelity. ​
• Second, building quantum circuits that remain efficient, accurate and interpretable within a constrained number of qubits while minimizing sensitivity to quantum noise. 
• Third, ensuring the model maintained its classification performance despite conversions and downscaling needed for quantum compatibility.  ​

All challenges required tight integration between quantum design, preprocessing, and model optimization.

Solution

We built two quantum models — a Quantum Neural Network (QNN) and a Quantum Convolutional Neural Network (QCNN) using IBM’s Qiskit platform. To prepare the data, we first converted RGB images to grayscale and resized them for consistency. Then, we used quantum encoding to turn image pixels into quantum states.​

The QNN model used 16 qubits and 32 trainable parameters, with Quantum gates helping it learn to classify images into categories. The QCNN followed a structure similar to classical CNNs and used quantum gates to create entangled “convolution” layers. Both models were trained and tested on unseen images, and we used the Adam optimizer to find the best balance between accuracy and efficient use of qubits.

Impact

The final models achieved up to 96% classification accuracy, demonstrating parity with classical models while consuming far fewer computational resources. This approach offers significant value to governments, construction firms, and asset managers by enabling cost-effective, real-time defect detection in field-deployable edge environments. The compact quantum architecture supports autonomous infrastructure monitoring and opens new avenues for integrating quantum AI into predictive maintenance workflows.

Services Rendered

• Image Preprocessing & Grayscale Normalization​
• Quantum Angle Encoding Integration​
• Quantum Circuit Optimization​
• Model Compression within Qubit Constraints​
• Unseen Data Evaluation​
• Quantum-Classical Comparative Visualization