High-Precision Mode Decomposition With Low Data Requirements Using Deep Learning

In this paper, we propose AttenResNet-MD, a method that combines channel attention mechanisms with a residual network architecture, enhancing feature learning efficiency and significantly reducing both the required training dataset size and training time. In simulation tests, AttenResNet-MD achieves high-precision mode decomposition with a modal weight error of less than 0.23% and a modal relative phase error of under 0.39% in the 3-mode case using only a training set of 10,000 beam patterns. Notably, this method achieves advanced mode decomposition accuracy while reducing the training set size by at most 10 times compared to conventional approaches. In experimental tests, the reconstructed beam patterns exhibit a high similarity of over 98% with the real beam patterns captured by a CCD.