<![CDATA[Monitoring digital television broadcasting stations requires accurate extraction of technical measurement parameters from screenshot images generated by television analyzer devices. In current practice, parameters such as frequency, signal level, bandwidth, modulation, guard interval, code rate, Fast Fourier Transform (FFT), pilot pattern, provider name, and video standard are often read and recorded manually, which may cause transcription errors when the images contain small characters, low resolution, noise, or visually similar symbols. This study proposes an Optical Character Recognition (OCR) system based on a hybrid ResNet18-Bidirectional Long Short-Term Memory-Connectionist Temporal Classification (ResNet18-BLSTM-CTC) architecture to improve technical parameter extraction from digital television analyzer screenshots. ResNet18 is used to extract visual character features, BLSTM models sequential relationships from both directions, and CTC decodes character sequences without explicit character segmentation. The implementation includes image acquisition, Region of Interest (ROI) selection, cropping, grayscale conversion, resizing into 48 x 48 pixel character images, model training, fine-tuning using hard labels, comparative evaluation, and web-based deployment using Streamlit. After fine-tuning, the proposed model achieved 96.58% accuracy, 0.0342 Character Error Rate (CER), 0.9253 macro F1-score, and reduced errors from 80 to 55. Additional comparison with Tesseract, EasyOCR, and CRNN-CTC, as well as testing on a different analyzer dataset, shows the practical relevance of the proposed OCR workflow for digital television technical monitoring.]]>
Copyrights © 2026
