The Autocorrelation Function Denoising Algorithm Based on Convolutional Neural Networks

To enhance the image reconstruction quality of the Nth-order linear (NL) linear algorithm, this study proposes a convolutional neural network (CNN)-based denoising method that processes noisy electric field temporal autocorrelation functions <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</i>1(τ) prior to reconstruction. The lightweight CNN architecture effectively suppresses noise by learning the mapping between noise-free and noisy autocorrelation functions, significantly improving the reconstruction quality when using noise-corrupted <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">g</i>1(τ) data.Evaluation of before and after denoising reconstruction demonstrates the network's superior noise reduction capability. Simulation tests reveal denoised root mean square errors (<italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">RMSE</i>) of 0.169, 0.105, 0.195, and 0.138 for four distinct anomaly types. Phantom experiments show <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">CONTRAST</i> values of 0.736 and 0.844 for quasi-solid cross-shaped and bar-shaped anomalies after denoising, while liquid tubular anomalies with varying flow velocities exhibit consistently higher contrast compared to noisy reconstructions. The deep learning approach achieves rapid and substantial noise reduction in autocorrelation curves, leading to markedly improved image reconstruction quality.