Hand-Picked Text-Focused DocLayNet Subset for Document Super-Resolution Tasks

About Dataset This dataset is based-on the DocLayNet dataset, specifically hand-picked to support research on document restoration particularly in cases involving textual documents. The selection process involved identifying and isolating images that predominantly contain text, ensuring that the model focuses on learning features relevant for text restoration and enhancement. The dataset is partitioned into: Training Set: 1000 images Validation Set: 200 images Test Set: 200 images Each image in the dataset is provided as a High Resolution (HR) as ground truth and Low Resolution (LR) pair, suitable for supervised learning approaches. This dataset was employed for the research work titled:"Beyond OCR: GAN-Driven Restoration of Severely Degrading Document" Experimental Results Summary The image enhancement process using this dataset demonstrated a peak Character Error Rate (CER) reduction of 3.20% compared to the low-resolution input. Quantitative evaluation using CER highlights this improvement, especially around epoch 2000 to 3000. Interestingly, epoch 2000 achieved the best CER improvement (−3.20%), yet its PSNR score was lower than that of epoch 3000. On the other hand, epoch 1500 recorded the highest Peak-Signal to Noise Ratio (PSNR) score at 20.7596 dB, but counterintuitively resulted in a 4.45% increase in CER, suggesting overfitting to perceptual factor that don't enhance actual text recovery. This reveal that both epoch 1500 and 2000 suffered from local optima, which limited generalizability and showing model instability. After further training to epoch 3000, the model achieved a balanced performance, reducing CER by 2.52% while also demonstrating improved perceptual quality, both at a globallayout level and in fine-grained text restoration. This indicates a more robust generalization, showing that the model had learned to balance between perceptal reconstruction and text readability. Epoch SSIM PSNR (dB) Character Error Rate (Before Restoration) Character Error Rate (After Restoration) 500 0.9987 18.7948 0.9847 2.1999 1000 0.9982 18.3975 0.9851 1.9996 1500 0.9994 20.7596 0.9855 1.0300 2000 0.9984 18.4169 0.9846 0.9526 2500 0.9989 18.6938 0.9845 1.2209 3000 0.9993 20.6035 0.9848 0.9596