Image Denoising with Hybrid Classical-Quantum Convolutional Neural Network

Noises can be introduced during the image acquisition process due to the factors such as electronic device malfunctions, making denoising essential for enhancing image quality and interpretability. However, existing denoising neural networks often struggle with generalization when training data is limited, resulting in suboptimal performance. To address these challenges, a novel quantum denoising convolutional neural network (QDnCNN) is proposed to enhance image denoising effectiveness. This approach integrates quantum circuits prior to the input of a classical convolutional network to extract features from noisy images in quantum space. The processed images are then rendered as the input of the convolutional neural network, significantly improving denoising performance. The quantum layers enable the network to effectively extract and represent image information in a higher-dimensional feature space, capturing more noises while preserving finer image details, thereby enhancing robustness to noise. Additionally, a quantum-parameterized convolutional layer is introduced to perform local transformations on image data, facilitating more efficient feature extraction. Experimental results demonstrate that the QDnCNN outperforms traditional networks under the Set12, BSD68, and MNIST datasets. Under the real-world datasets Set12 and BSD68, the QDnCNN achieves performance comparable to the conventional deep learning methods, while under the MNIST dataset, it improves the structural similarity index by 11.4% and the peak signal-to-noise ratio by 7.6%. The QDnCNN shows excellent potential as an image denoising algorithm suitable for future quantum computing devices.