Auto focusing of in-Line Holography based on Stacked Auto Encoder with Sparse Bayesian Regression and Compressive Sensing

In recent years, Digital holography has emerged as an exceptional imaging technology for tracking high-contrast object particles and, interestingly, analyzing 3D object data in real time. The best quality images can be obtained effectively using the auto-focusing algorithm. In this paper, the focus location of the object is traced with a deep learning-based auto-focusing algorithm. The proposed model constructs a large feature pool by considering different focus measures to reconstruct objects from two out-of-focus images. The preferred features are selected through the proposed Support vector Machine-based Recursive Feature Elimination (SVM-RFE) method. Therefore, the inappropriate features are eliminated, and the reconstruction distance is obtained by the suggested stacked autoencoder with sparse Bayesian regression (SAE-SBR) model training. It is common to find a twin image in the reconstructed image, and such noise interference is minimized with the presented high-speed iterative shrinkage/thresholding (HS-IST) based compressive sensing (CS) algorithm. Reconstruction distances are predicted by the proposed method with a standard variation of about 0.036 mu m. The proposed SAE-SBR predicts the right reconstruction distance of a single hologram, and it is 600 times faster than traditional autofocusing techniques like Dubois and Tamura of Gradient (ToG). Also, the computation time of the proposed model is 33.3% less than the well-known FocusNET model.