Skip to content
DSML.KZ
Publications

Talk · International Symposium on Semiconductor and Electronic Technology · July 1, 2026

Calibrated Degradation-Aware Augmentation for Robust Complex Street-View Text Detection

Jinghui Ma, A.Dzh. Kartanova

Asel Kartanova
Open via DOI

Abstract

Street-view text detection is sensitive to small text and capture degradation. This paper studies a compact detection-only pipeline combining high-resolution one-stage detectors with Calibrated Degradation-Aware Augmentation (CDAA). On ICDAR2015 validation, YOLOv8s improves from 0.7415 mAP50 at 640 to 0.8097 at 1152. Medium CDAA provides the clearest degraded-set robustness for YOLOv8s, while YOLO11s with curriculum CDAA is selected as the balanced final model. Supplementary MSRA-TD500 evidence is mixed because oriented text is simplified into horizontal boxes.

Handwritten Text Recognition TechniquesAdvanced Neural Network ApplicationsVehicle License Plate Recognitionone-stage object detectionYOLOv8curriculum learning

More by Asel Kartanova