Speckle noise reduction in coherent imaging based on deep learning without clean data

In this paper, we propose a new algorithm based on deep learning to reduce the speckle noise for coherent imaging without clean data. By learning the common information, namely the clean image, from paired noisy holographic reconstructed images, the noise reduction mechanism of the system can be obtained. Unlike normal deep learning methods, our algorithm does not require prior knowledge of clean object distribution. Experimental results show that noise removal effect of proposed method is better than traditional smoothing algorithms and can be comparable to the existed deep learning method trained with clean data.