Add the PARSeq model for Scene Text Recognition

[gowthamkpr]

Adds the PARSeq OCR model for scene text recognition (STR): https://arxiv.org/abs/2207.06966

Features:

• iterative autoregressive and non-autoregressive decoding and refinement, fully performed within Keras' computation graph
• training based on permutation sequences, as described in the paper
• ViT-based recognition model
• well tested model with state-of-the-art performance

This is (so far) a 1:1 translation of PaddleOCR's implementation (Apache 2.0-licensed), in particular of

• https://github.com/PaddlePaddle/PaddleOCR/blob/main/ppocr/modeling/backbones/rec_vit_parseq.py
• https://github.com/PaddlePaddle/PaddleOCR/blob/main/ppocr/modeling/heads/rec_parseq_head.py
• https://github.com/PaddlePaddle/PaddleOCR/blob/main/ppocr/postprocess/rec_postprocess.py#L586
• https://github.com/PaddlePaddle/PaddleOCR/blob/main/configs/rec/rec_vit_parseq.yml

Next steps:

• Port PaddleOCR weights
• Adaptations to meet KerasHub style
• Pre/postprocessing, and a dedicated Task for OCR
• Add tests

[mattdangerw]

Thanks! When this is done, could you attach a colab showing basic usage? Or even just a code snippet that shows the full usage including data input formats would be helpful.

[gowthamkpr]

A few comments on code layout of this first version of the code:

• There's a weird bug where tests for all backends complete perfectly fine when run on their own, but fail when run after other tests. This occurs because Keras rewrites the class hierarchy of Backbone as soon as another model is instantiated (see here) with __init__(inputs=..., outputs=...), therefore the expected __init__ arguments change and inputs and outputs arguments always have to be provided.
  If this is a problem with a known solution, please let me know. I'll investigate further. Due to the way inference and training work with PARSeq, we require a call method here.

Document explaining where the bug happens!!!

• Compared to the existing ViT implementation, the ViT implementation I've adopted from PaddleOCR contains some additional logic for Drop Path regularisation, but the pretrained weights do not appear to use this. For being able to use the pretrained weights alone, we might alternatively just use the existing ViT implementation, if that is preferred.

• For now, I've implemented a simple detokenize method. Alternatively, we might add a full-fledged Tokenizer to support the model character set, either extending the existing ByteTokenizer to support a given character set, or creating a new one.