CHARACTER RECOGNITION IN LOW RESOLUTION IMAGES USING ALTERNATIVE COLLABORATIVE LEARNING

Low-resolution (LR) images pose a significant challenge for accurate character recognition, especially in realworld applications such as license plate identification and traffic surveillance. This paper presents a deeplearning framework inspired by Alternative Collaborative Learning (ACL), designed to enhance recognitionperformance in degraded visual conditions. The proposed method integrates a Deep Back-Projection Network(DBPN) for image super-resolution with a convolutional OCR model, trained using a two-phase alternativecollaborative strategy. Specifically, during training, the SR and OCR networks are alternately optimized whilethe other is frozen, enabling stable convergence and mutual reinforcement. A custom dataset of 11,000 highresolution license plate images was used, where HR images (180×64) were downscaled by a factor of 4 togenerate LR inputs (45×16). The model was evaluated using Peak Signal-to-Noise Ratio (PSNR) and meanAverage Precision (mAP) metrics. For a scale factor of ×4, The ACL framework outperformed traditional superresolution pipelines and baseline OCR models. This demonstrates the effectiveness of collaborative learning forrobust character recognition in low-quality images.