A squeeze-and-excitation-attention-block-based SAR denosing network

As an active observing system, synthetic aperture radar (SAR) is often used in remote sensing. It is widely applied in civil and military since it can do all-weather, all-day, and all-night imaging. Typical applications include ocean monitoring and ship classification detection. However, SAR images have coherent speckle noises, and those noises significantly damage the image quality and the following interpretation. To effectively reduce speckle noises in the SAR images and preserve the image details, we put forward a squeeze excitation attention based SAR denoising network (SEADNet). In this paper, we aim to break the limits of the traditional methods and balance the denoising and detail preservation. Specifically, we first build a feature denoising block (FDB). It uses average pooling to denoise the feature map, which enhances the network’s anti-interference ability without adding additional parameters. Furthermore, we also designed a channel and spatial attention block (CSAB) and a squeeze and excitation attention block (SEAB). Those blocks use dilated convolution networks and global average pooling to enhance the network’s attention to areas with rich information. This design enables the model to retain essential information and fine structural details within the image. To evaluate the effectiveness of the proposed approach, experiments were conducted on real SAR image datasets, with performance compared against both traditional and state-of-the-art methods. Experimental results show that the proposed method achieved an Equivalent Number of Looks (ENL) of 6.0422, and Edge Preservation Degree based on the Ratio of Averages (EPD-ROA) values of 0.9564 and 0.9212 in the horizontal and vertical directions, respectively. These results indicate that the proposed SEADNet offers substantial improvements in both visual quality and quantitative metrics over existing SAR image denoising techniques. Moreover, we effectively reduce the network computational complexity. This method can be generalized in many areas and has great practical value.