I am working on a Undergraduate Thesis for my Information Systems Bachelors degree and I have proposed building a software to extract "burned-in" subtitles from videos and saving them to .srt format with the correct timestamp. I am still working on the theoretical foundation for my thesis but I'm already thinking about what tool/libs I'm going to use for the software itself. In theory, the project is supposed to use image processing, ocr and deep learning techniques to do the job, but I'm unsure on what to use to build that in Python, as I have only had used OpenCV and Tesseract for other personal projects. The pipeline I'm thinking about is basically: Image pre-processing (resizing, etc) Subtitle area recognition (bounding box) Image processing techniques to help OCR Extract subtitles with OCR engine (Tesseract, ...) Somehow will sync all that with the subtitle's correct timestamp I've seen there are lots of ways of detecting/recognising and extracting text from images, like using deep learning techniques to find the text area/extract it (read about lots of cnn architectures but its all new to me) and just traditional image processing + ocr. I've seen MMOCR pretty much does all the job I need but I was thinking maybe I wouldnt have such a complex enough software to present for my thesis then. Therefore, I was thinking of mixing different tools to achieve each step of the pipeline I presented above while trying to mix deep learning, image processing and OCR to my project, as I have stated to my professor I would do. I would accept any tips on libs/tools so I could start from somewhere, thank you all very much. Continue reading...
