Data Equalization Distribution Improves the Near-infrared Tissue Reconstruction based on Stacked Auto-encoder

The near-infrared optical imaging technology based on deep learning has attracted much attention in the field of imaging reconstruction due to its small amount of calculation, fast reconstruction speed and so on. Modeling sample datasets selection are directly related to the accuracy and stability of the training model. Aiming at the influence of randomly selecting data samples on the effect of optical reconstruction based on deep learning, this paper proposes a method for selecting data samples based on equal distance cross-selection to achieve data equalization distribution. Based on the stacked auto-encoder neural network, the imaging model of 350 data samples was established, and the remaining 80 data samples were predicted. The results show that the prediction accuracy of anomaly reconstruction is 77.2% under the method of randomly selection sample datasets, while the training datasets and the prediction datasets were processed by the data equalization distribution selection method, the SAE method achieved the prediction accuracy of anomaly reconstruction of 96.25%. The method of data equalization distribution selection to collect modeling sample datasets can not only improve the accuracy of optical imaging detection effectively, but also have a certain guiding significance for the selection method of optical reconstruction sample datasets based on deep learning.