SAR-NODE: SAR DESPECKLING USING NEURAL ORDINARY DIFFERENTIAL EQUATIONS

Synthetic Aperture Radar (SAR) images have higher resolution and are widely used in practice. However, they are easily contaminated by speckle noise. While deeper networks have proven effective for despeckling, their increased number of learnable parameters can lead to over-fitting and necessitate larger datasets, which is problematic since SAR datasets are difficult to obtain. Additionally, existing deep learning models frequently lack interpretability, making their outputs unpredictable. To address these challenges, we propose the SAR-NODE despeckling model based on the neural ordinary differential equation (NODE) framework and wavelet transform. Firstly, the input image is transformed into the frequency domain to separate noise from image features. We then design two specialized denoising blocks using the NODE framework and the attention mechanism to process the low-frequency and high-frequency components separately, while also enhancing the model’s interpretability. Finally, an inverse transform is applied to recover the image to its original domain without any loss of information. Experiments on both synthetic and real SAR images demonstrate the superior performance of the SAR-NODE model, both quantitatively and qualitatively.