chaochun zhong and Qingliang Ma
Adv. Artif. Intell. Mach. Learn., 5 (4):4444-4460
1. chaochun zhong: Guangzhou Institute of Measurement And Testing Technology
2. Qingliang Ma: Guangzhou Institute of Measurement And Testing Technology
DOI: 10.54364/AAIML.2025.54247
Article History: Received on: 26-Jul-25, Accepted on: 23-Oct-25, Published on: 30-Oct-25
Corresponding Author: chaochun zhong
Email: bellamyroache018@outlook.com
Citation: Qingliang Ma, et al. Enhancing OCR Performance Through Super-Resolution Reconstruction Using SSAE-REAL-ESRGAN. Advances in Artificial Intelligence and Machine Learning. 2025;5(4):247. https://dx.doi.org/10.54364/AAIML.2025.54247
In OCR tasks, the blurring and noise of low-quality text images often affect the accuracy of character recognition. To improve the performance of OCR systems in complex scenes, this paper proposes a super-resolution reconstruction method that combines a Stacked Sparse Autoencoder (SSAE) with Real-ESRGAN. This method leverages the hierarchical feature learning and sparse constraint mechanism of SSAE to extract key text information and suppress background noise, while utilizing the high-quality reconstruction capability of Real-ESRGAN to enhance the recovery of character edges and structural details. Experimental results on the Text-Zoom dataset show that this method outperforms the ESRGAN model by 10.27% in terms of PSNR and 10.38% in terms of SSIM, effectively improving the OCR system's ability to recognize low-quality text images.