PCA neural network for JPEG image enhancement

Principal component analysis (PCA) is a statistical method capable of transforming multivariate data into components that are statistically uncorrelated from each other. We employ the adaptive principle component extraction-based nonlinear PCA to extract the principal components from the transform encoded images and apply the neural network based regularization with the extracted principle components from the images for image enhancement. Unlike other regularization approaches, the PCA-based regularization requires no a prior knowledge and thus no extra bits are added for the coding overhead. The experiment on JPEG fingerprint images and on wavelet (SPHIT) encoded images has shown that with the well-trained PCA and neural networks, the degraded images can be enhanced by 15-20% in the PSNR measurement.