基于对抗机制与正交约束的自编码器图像重建OA

The Principal Orthogonal Latent Component Analysis Network(POLCA Net)was an autoencoder-based extension method for structured representation learning.By integrating classification supervision,orthogonality constraints,and variance constraints,it achieved ordered and mutually orthogonal latent components,thus demonstrating remarkable advantages in dimensionality reduction and feature extraction.However,due to the deterministic encoding process of POLCA Net,it could not effectively model the uncertainty of data distribution.Moreover,its joint loss function inhibited the latent space's ability to express high-frequency information,ultimately making it difficult to ensure high reconstruction quality in image reconstruction tasks.This study proposed an autoencoder algorithm and framework integrating adversarial training and orthogonality constraints,named GAN-POLCA(Generative Adversarial Network in Principal Orthogonal Latent Components Analysis)to address these issues.While retaining all the original constraint mechanisms of POLCA Net,GAN-POLCA introduced an adversarial loss and constructs a multi-constraint collaborative deep generative model through the adversarial training mechanism between the generator and discriminator.Experimental results on three real datasets showed that GAN-POLCA significantly outperformed PCA(Principal Components Analysis)and POLCA Net in both the convergence rate and steady-state value of the loss function,with effectively improved image reconstruction performance,which fully verified the superiority of the proposed algorithm and framework.