Self-referencing spectral constraint-based spectral reconstruction method and apparatus, device, and medium

A self-referencing spectral constraint-based spectral reconstruction apparatus, comprising:

• a calibration module, configured to perform a calibration to acquire a total transmission matrix;
• a measurement module, configured to acquire an ith observed speckle generated after target light passes through an ith filtering interval of a spectral filter to enter a scattering medium, wherein i=1, 2, . . . , K;
• K filtering intervals correspond to K observed speckles;
• and the K observed speckles form a total observed speckle;
• a self-referencing spectrum calculation module, configured to sum an intensity of the ith observed speckle, and acquire a total intensity of the ith observed speckle, wherein total intensities of the K observed speckles form a self-referencing spectrum;
• a network optimization module, configured to:
• reconstruct a total target spectrum;
• construct a spectral reconstruction model based on a neural network and a physical forward process of the total transmission matrix;
• construct a composite loss function based on a speckle intensity estimation error and a self-referencing spectrum error;
• and optimize parameters of the neural network via the composite loss function until convergence, yielding an optimized neural network for a spectral reconstruction, wherein
• the speckle intensity estimation error is acquired by:
• calculating, based on a physical model F(⋅) for the total observed speckle, the total transmission matrix and the total target spectrum, a total observed speckle estimate corresponding to a total target spectrum estimate output by the neural network;
• and calculating a mean squared error between the total observed speckle estimate and
• the total observed speckle, and acquiring the speckle intensity estimation error ILoss;
• and
• the self-referencing spectrum error is acquired by:
• inputting the total observed speckle into the neural network, outputting the total target spectrum estimate, downsampling the total target spectrum estimate, calculating a mean squared error between the total target spectrum estimate and the self-referencing spectrum, and acquiring the self-referencing spectrum error SLoss;
• and
• a spectral reconstruction module, configured to perform the spectral reconstruction based on the optimized neural network and the total observed speckle.