A Two-Component Deep Learning Network for SAR Image Denoising

Speckle noise reduction and detail preservation, between which a balance is hard to achieve, are two main purposes of synthetic aperture radar (SAR) image denoising. For different regional characteristics of SAR images, satisfactory denoising results are obtained by complex parameter fine-tuning in most methods, and a solution with strong robustness seems difficult to find. In this paper, a novel two-component deep learning (DL) network is proposed to solve the above problem. First, the texture estimation subnetwork is constructed to produce the texture level map (TLM), which evaluates the randomness and scale of the texture distribution. Then, the noise removal subnetwork learns a spatially variable mapping between the noise and clean images with the help of TLM. Once the network has been trained, it can automatically quantify the texture feature and decide whether to smooth the local noise or maintain the detail. Comprehensive experiments on simulated and real SAR images demonstrate the superior performance of the proposed method over the state-of-the-art methods with respect to both the visual effect and quantitative analysis.