Topic: Impact of Image Compression on Machine Handwriting Recognition
- you will investigate the impact of lossy image compression algorithms, particularly JPEG
images, on the accuracy of machine handwriting recognition systems.
- TrOCR (Transformer-based Optical Character Recognition),
- Dataset : we use handwritten images from the IAM Handwriting Database with different
compression levels as input for the model. Each level is assigned a specific percentage of image
quality, for example, images with compression levels of 30%, 40%, 50%, 60%, and so on.
- Idea for the dataset: The dataset will be the IAM Handwriting Database, and the images should
be in JPEG format. - We need to write a script to compress each image to 20% quality and save it
in a file with the name: line_(imagename)_20. - We will repeat the process for other
compression levels, such as 30%, and save them as: line_(imagename)_30. - The same process
will be applied for all compression levels (e.g., 40%, 50%, 60%, etc.).
- ML Modell it must be TRORC
- Encoder
- Decoder https://huggingface.co/docs/transformers/model_doc/trocr
We trained the model with images compressed at 20% and at all compression levels to
demonstrate the impact of JPEG compression on our TRORC model.
- Modellevaluation
• CER (Character Error Rate):
• WER (Word Error Rate):
-measure to improve the modell
-Training with highly compressed images.
-Using a noise reduction algorithm as a preprocessing step
