Development
The system became useful one failure at a time.
This is not a corporate timeline. It is a record of how each observed failure created the need for another part of the system.
Image to text
A basic OCR pass proved that printed characters could become machine-readable, but it also exposed how strongly recognition depended on the image.
Live camera and speech
Camera input and text-to-speech turned the experiment into a reading loop. They also showed that the user needed feedback before and after recognition.
Segmentation and quality checks
Page detection, text-region localization, blur detection, and glare detection moved the project from hopeful OCR calls toward deliberate input control.
Custom recognition and fallback
A CRNN-CTC model provides a research path while Tesseract remains a practical fallback. Candidate selection makes disagreement visible rather than pretending one engine is always right.
Burst stacking and edge optimization
Median frame stacking addresses temporary variation, while Raspberry Pi constraints force decisions about frame skipping, depth, exposure, and service boundaries.
Adaptive spoken guidance
The latest system direction closes the loop: detect the problem, speak one useful adjustment, and capture again when the page is ready.