Multimode fiber (MMF) plays a vital role in promoting the miniaturization of endoscope. However, real-time and high-definition imaging using the MMF that remains a challenging research. Traditional phase compensation and transmission matrix methods are affected by fiber shape and optical devices, which results in low imaging rate and accuracy. Deep learning can be used to construct the inverse transformation matrix (ITM, output to input) of the MMF. However, deep learning requires high similarity between sample sets. In this paper, we combine principal component analysis (PCA) method, deep learning based speckle classification (DLSC) and deep learning based image enhancement (DLIE) to improve imaging definition. To save experimental costs, we use the inverse-PCA method to obtain simulation speckles. The experimental results show that simulation speckles can be used for classification and image reconstruction of experimental speckles. With the difference between simulation and experimental speckles, which brings about low imaging definition. Therefore, we use the DLIE methods to further improve imaging definition. The experimental results show imaging capability with high definition for complex natural scenes, which may provide a feasible method for high definition images transmission through the MMF.

多模光纤（MMF）在促进内窥镜的小型化方面起着至关重要的作用。但是，使用MMF进行实时和高清成像仍然是一项具有挑战性的研究。传统的相位补偿和传输矩阵方法受光纤形状和光学器件的影响，导致成像速率和准确性较低。深度学习可用于构造MMF的逆变换矩阵（ITM，输出到输入）。但是，深度学习需要样本集之间的高度相似性。在本文中，我们将主成分分析（PCA）方法，基于深度学习的斑点分类（DLSC）和基于深度学习的图像增强（DLIE）相结合，以改善成像清晰度。为了节省实验成本，我们使用逆PCA方法获得仿真斑点。实验结果表明，仿真斑点可用于实验斑点的分类和图像重建。由于仿真斑点和实验斑点之间存在差异，因此导致成像清晰度较低。因此，我们使用DLIE方法来进一步改善成像清晰度。实验结果表明，对于复杂的自然场景具有高清成像能力，这可能为通过MMF传输高清图像提供一种可行的方法。