This invention relates to the field of computational
optical imaging and
microscopic image processing technology, and discloses a method and
system for
joint reconstruction and denoising of
light field microscopy based on self-
supervised learning. In the pre-training stage, the method performs random masking operations in the spatial and angular domains on large-scale
light field data, guiding the
encoder to learn general features through a completion task. In the model training stage, the acquired
light field images are randomly divided into input and supervision
viewpoints. Features are extracted using the
encoder and aligned in three-dimensional space using a
physical point spread function. A three-dimensional volume is generated through viewpoint
pooling and a decoding network, and a self-supervised loss is established using the supervision viewpoint, achieving implicit denoising based on the independence of
noise distribution between
viewpoints. In the testing stage, high
signal-to-
noise ratio three-dimensional reconstruction results can be obtained by inputting all
viewpoints. This invention does not rely on
noise-free
ground truth images, can adapt to different optical
system configurations, and improves the reconstruction quality and denoising capability of
light field microscopy imaging.