Multi-Scale Feedback Convolutional Sparse Coding Network for Saliency Detection in Remote Sensing Images

Due to the huge difference in the shooting conditions of remote sensing images (RSI), the RSI itself has variable scales, multiple scenes and cluttered backgrounds, which lead to the poor detection effect of the SOD method for natural scene images (NSI). By exploring the multi-scale characteristics of RSI, combining convolutional sparse coding (CSC) and convolutional neural network (CNN), this paper proposes a multiscale feedback CSC (MFC) network for SOD of optical RSI. Specifically, the soft threshold shrinkage (SST) function and the CNN components are first used to construct the CSC block (CSCB). Then, the multi-scale image representations are fed into the stacked CSCB to extract the features thoroughly. Finally, the side-out features are integrated through the cross-feature fusion module (CFF) with a top-down feedback strategy. The comprehensive evaluation results with the state-of-the-art (SOTA) competitors on the ORSSD and EORSSD datasets demonstrate the proposed model's priority.