ACNet: a twin-image removal network based on CNN with attention mechanism in in-line lensless digital holographic microscopy

Transparent objects are extensively utilized across various aspects, yet their non-destructive optical measurement remains challenging. In-line lensless digital holography has emerged as an efficient and precise technique for detecting transparent objects, with the advantages of simpler device requirements and more effective utilization of the detector limited spacebandwidth product. However, the presence of twin-image significantly degrades the quality of the reconstructed images. Conventional approaches to mitigating twin-image require intricate hardware configurations or time-consuming algorithms. In this paper, we proposed a new network called Attention mechanism in Convolutional neural Network (ACNet), which provides a fast and efficient deep learning solution for twin image suppression. The proposed approach utilized numerically generated datasets for training and a convolutional neural network (CNN) was employed with an attention mechanism to perform twin image removal. Simulation results demonstrate that this method effectively eliminates twin image interference in phase recovery, thereby enhances the reconstruction quality of in-line digital holography. The present work has great potentials for wider applications in digital holography.