Research on the reconstruction method of two adjacent object speckle images based on DCGAN

Abstract In recent years, deep learning has been successfully applied to the reconstruction of speckle images formed through scattering media. However, most research on imaging through scattering media has mainly focused on reconstructing images of a single object, where the object’s information can be extracted from a single speckle using convolutional neural networks. Reconstructing images of multiple objects from a single speckle is more important and challenging, as the information of the objects becomes highly mixed during the light propagation process. Moreover, in the speckle imaging process, most neural networks are trained to predict pixel-by-pixel, treating each pixel of the image independently. This can lead to a lack of spatial continuity in the final result. To achieve better performance, the network should not only focus on the class features of each pixel value but also consider enhancing the visual appearance of the reconstructed image. In this paper, a CNN and GAN-based network model is designed for speckle image reconstruction. The model consists of an encoder and two decoders. A network called DCGAN (Double_CNN_GAN) is proposed to reconstruct speckle patterns. By using DCGAN, we achieve high-fidelity simultaneous reconstruction of two different binary or grayscale object images located behind the scattering medium. Additionally, the influence of the distance between the two objects on the reconstruction quality is explored. Therefore, the study of methods for reconstructing the speckle images of two adjacent objects is of significant theoretical and practical importance.