‘Pre-training+Fine-tuning’ Model-based 1/2 Folded Image Restoration

The 1/2 folded image refers to folding the original image at the halfway point along the horizontal or vertical direction, creating an image with partial overlap. The purpose of restoring 1/2 folded images is to repair folded, damaged, or partially hidden images in order to restore image integrity, recover data, analyze and process images, and perform reverse hiding and steganography analysis. In response to the restoration problem of 1/2 folded images, a model called ‘pre-training + fine-tuning’ was applied to restore folded images using unsupervised pre-training and supervised fine-tuning modes. The WGAN-GP model with gradient penalty is used for the pre-training model. A four-layer non-linear fully connected network is selected for the fine-tuning model. The experimental process consists of two stages. Firstly, the WGAN-GP network is subjected to unsupervised pre-training to obtain a generator capable of generating clear images. Then, in the fine-tuning stage, the output of the generator is used as the input for a fully connected network, and the target results of the fully connected network model are fixed for fine-tuning, aiming to restore the folded images to be close to the original ones. The applied method is validated on the MNIST and Fashion-MNIST datasets in this paper. The experimental result shows that the “pre-training + fine-tuning” mode achieves good restoration performance on 1/2 folded handwritten digit images and Fashion-MNIST images, with the maximum similarity score between the restored images and the original images is around 0.98. Moreover, this technique can enhance image quality and information integrity, expand the application areas of image processing and computer vision.