ACGNet: An Alternating Conjugate Gradient Optimization-Based Neural Network for SAR Image Despeckling

Synthetic Aperture Radar (SAR) images are characterized by unique speckle noise, and maintaining image details while effectively reducing this noise has always been a challenging problem. Current deep learning-based denoising methods mainly rely on global noise models or local feature learning, but these methods often fail to achieve a balance between noise suppression and detail preservation. To address this issue, this paper proposes a supervised collaborative denoising method with alternating optimization, which combines the alternating conjugate gradient method with a SAR despeckling network trained on paired noisy-clean simulated SAR data to progressively optimize image quality, thereby effectively reducing noise while preserving more details and texture information in the image. Additionally, to enhance the feature learning capability of the denoising network, a DEGA (Detailed Prior Enhancement Convolution and Gated Weighted Attention) module is introduced. This module enhances feature learning capability and improves image detail extraction by guiding network decision-making through adaptive weights. To better preserve the texture details of the image, an edge correction function is designed, incorporating edge correction terms and regularization terms into the traditional Mean Squared Error (MSE) loss function, thereby effectively improving the performance of the denoising results in edge regions. The proposed supervised framework is trained end-to-end on simulated SAR datasets with ground-truth references, and demonstrates exceptional performance on both simulated datasets and real SAR images, as evidenced by extensive experimental results. Furthermore, comparative analysis highlights the substantial advantages of the method in both visual quality and quantitative metrics over classical and modern despeckling techniques. The implementation code for this study is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/4mx12/ACGNet.git</uri>.