A two-stage enhancement network with optimized effective receptive field for speckle image reconstruction

Reconstructing target objects from strong speckle images is a key step for solving complex inverse scattering imaging problems. Deep learning (DL) methods are very effective for producing high quality object reconstruction, especially for speckle image reconstruction (SIR). Understanding the relationship between DL network structures and reconstruction results helps improve the reconstruction quality. Although previous studies have explored this issue, few of them considered dilated convolution adjustment and effective receptive field optimization of DL networks in image reconstruction for improving the reconstruction quality. In this paper, we propose a two stage enhancement network for speckle image reconstruction, in addition, we present an effective receptive field optimization method for maximizing the usage of the network capability. Specifically, in the first stage, we propose a growth model exploiting the dilation rates under the assumption that the central area pixels of images have a much bigger impact on the output field than the outer area pixels, and accordingly optimize the effective receptive field of the networks. Then, based on our growth model, in the second stage, the enhancement network jointly utilizes complementary information from the objective loss and perceptual loss when reconstructing objects. Extensive experiments show that our new network outperforms five state-of-the-art methods in the MAE, MSE, PSNR, and SSIM evaluating measures.