Inferring MAXIM for Text Deblurring: Enhancing Clarity in Image Restoration
Roger Julianto Angryawan, Alfi Yusrotis Zakiyyah, Kelvin Asclepius Minor · 2025
The restoration of blurred text images presents unique challenges, as it requires precise reconstruction of character shapes to ensure readability. This study evaluates the Inferring capabilities of the Multi-Axis Multi-Layer Perceptron (MAXIM) model, a versatile architecture for image restoration, on the TextOCR dataset. While MAXIM's pretrained weights, derived from general-purpose deblurring tasks, were able to unblur portions of text, the overall performance—assessed using Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM) fell short of expectations for text-focused applications. Among the evaluated pre-trained models, the RealBlur_R weights yielded the best results with 22.17 PSNR and 0.707 SSIM. The findings emphasize the limitations of directly applying general-purpose pre-trained models to text-specific deblurring and highlight the need for dataset-specific fine-tuning. This research provides insights into the potential and constraints of MAXIM in addressing the unique challenges of text deblurring tasks and includes a comparative evaluation against the Restormer model.