Semantic Super-Resolution of Text Images via Self-Distillation

Abstract

This research develops an effective single-image super-resolution (SR) method that increases the resolution of scanned text or document images and improves their readability. To this end, we introduce a new semantic loss and propose a semantic SR method that guides an SR network to learn implicit text-specific semantic priors through self-distillation. Experiments on the enhanced deep SR (EDSR) model, one of the most popular SR networks, confirmed that semantic loss can contribute to further improving the quality of text SR images. Although the improvement varied depending on image resolution and dataset, the peak signal-to-noise ratio (PSNR) value was increased by up to 0.3 dB by introducing the semantic loss. The proposed method outperformed an existing semantic SR method.

Semantic Super-Resolution via Self-Distillation and Adversarial Learning

“…Inspired by the Park's method [5], the semantic priors are incorporated into the SR process via self-distillation. That is, we distill the category-specific semantic information from SR images of the same category (e.g., ''text'' or ''face'') during training an SR generator.…”

“…Therefore, it was successfully used as a regularization technique for matching the predictive distribution of the network between different samples of the same label [10]. Also, it was used for matching semantic features between different images of the same category and guiding the network to learn the semantic information [5]. We also use self-distillation for the semantic learning in this study.…”

“…Chen et al [29] proposed a blind SR method using a semantic-aware texture prior obtained by applying mini-batch K-means to feature vectors of HR images. Recently, to leverage semantic priors of text images, Park [5] proposed minimizing the semantic difference between text SR images using self-distillation. To this end, he proposed a semantic loss that represents semantic differences between images and is computed using the pretrained VGG network.…”

“…To this regard, semantic SR methods have attempted to incorporate contextual or semantic information of images belonging to the same category into the SR process, resulting in more accurate SR results without over-fitting. As the most recent study, Park [5] introduced a semantic loss to measure the semantic difference between text images and proposed a semantic SR method that minimizes the semantic loss via self-distillation. The method allowed existing SR models to be trained to produce better text SR images and the Enhanced Deep Super-Resolution (EDSR) model [6] trained using the method outperformed the generative adversarial network (GAN)-based semantic SR model.…”

Satellite image super-resolution reconstruction based on ACGAN and dual-channel dense residual network

“…And in Section 5, we present final conclusion and future work. [1] and Enhanced Deep Residual Networks (EDSR) were applied [2], as well as the SRDenseNet model constructed by the subsequent DenseNet structure [3]. Lai et al combined a series of structural features and combined with the image pyramid structure to propose a more complex Laplacian pyramid [4] image super-resolution structure to achieve multi-scale super-resolution reconstruction of images.…”