This study proposes a writing-quality-aware model selection framework for image-based text recognition to address the heterogeneity of real-world data. Three representative models, including CNN, ResNet combined with BiLSTM-CTC, and TrOCR, are evaluated across three data levels: printed text, clean handwriting, and poor handwriting. Experimental results show that model performance decreases as data quality deteriorates, while TrOCR consistently achieves the best robustness and accuracy. Based on these observations, the proposed framework consists of two stages: estimating the quality of input images and selecting the most suitable model for each quality level. Validation on independent datasets demonstrates that the framework maintains stable performance and improves overall accuracy compared to using a single fixed model, highlighting the effectiveness of quality-aware model selection for real-world OCR applications.