Neural Network-Based Intensity-Only Mode Decomposition With Angular Extraction

Mode decomposition based on neural networks (NN-MD) has the advantages of high accuracy, fast decomposition speed, and elimination of the need for iteration. However, existing NN-MD is affected by uncertainty and has insufficient decomposition capabilities in multi-mode scenarios. This problem mainly stems from the fact that traditional Cartesian coordinate system intensity distribution (CID) images do not have good enough denoising methods and information mining methods. In this paper, we proposed an angular intensity distribution (AID) method for NN-MD that focuses on capturing the inherent laws of the light patterns. Compared with the irregularity of the pixel intensity in CID, the pixel intensity of each row in the light pattern using AID can be expressed mathematically. Each row of pixels in AID is composed of trigonometric functions. This allows the AID-NN model to learn the implicit mathematical laws better. Moreover, our proposed method is compatible with other existing NN-MD algorithms by replacing CID with AID. This paper mentions a denoising method that cooperates with the AID method. The results of the CID method using an inception network that does not require pre-training are compared with the AID results using the same architecture. In simulations, AID-NN-MD exhibits smaller errors and lower standard deviations than CID-NN-MD across all modes. Furthermore, as the number of modes increases, the benefits of AID become more evident. In the highest mode number of simulation, CID-NN-MD's weight and phase errors are more than 50% higher than AID-NN-MD. In addition, experimental results also show ideal performance. Taking a pre-training-free network experiment containing six modes as an example, the cross-correlation factor between the captured image and the reconstructed image can reach 0.98.