Hierarchical YOLOv5 Detection and ResNet Recognition Pipeline for Degraded Heritage Character Imagery

Abstract

Character recognition on heavily degraded historical imagery presents three intertwined challenges: heterogeneous noise patterns generated by the underlying physical substrate, scarce labeled training data within the target domain, and small character instances embedded in cluttered backgrounds. This paper develops a four-stage pipeline that integrates adaptive image denoising, hierarchical YOLOv5 detection, multi-scale convolutional feature analysis, and ResNet recognition with cross-domain transfer learning. The denoising stage combines mean filtering with morphological opening to suppress impulse noise and blob-like artifacts, complemented by an attention-based generative adversarial network for residual texture artifacts. The detection stage trains YOLOv5 with initial learning rate 0.01, achieving test-set precision of 64.10% and recall of 49.50%, with the loss function stabilizing at 0.04 after sufficient iterations and outperforming YOLOv2, YOLOv3, Fast R-CNN, and R-CNN baselines on intersection-over-union, recall, and average precision metrics. Multi-scale feature visualization across convolutional output layers confirms that the detection model recovers character locations across 200 benchmark images and produces region coordinate vectors with high coverage. The recognition stage achieves over 70% accuracy under in-domain training, and incorporation of an external auxiliary dataset through transfer learning substantially improves recognition accuracy. Sensitivity analysis under additive noise confirms pipeline robustness across degradation levels.