Clarification Required on Running Malayalam Inference – IndicPhotoOCR

[dellajoseph1212-bit]

I have performed testing of the IndicPhotoOCR framework using this reference repository . As mentioned in the documentation, the model is expected to support multiple Indian languages, including Malayalam.

I have successfully:

1)Installed the virtual environment

2)Installed all required dependencies

3)Completed the setup process without errors

I attempted to test the following modules:

1)Detection

2)Cropped Word Script Identification

3)Cropped Word Recognition

4)End-to-End Scene Text Recognition

However, I am not receiving any output for Malayalam scripts. The execution completes without producing predicted text results.

Could you please guide me on the correct procedure for running inference specifically for Malayalam scripts? I would appreciate clarification on:

1)Whether any additional pretrained Malayalam-specific weights need to be downloaded

2)If any language configuration parameters must be explicitly set

3)Required input image format or preprocessing steps

4)Any additional checkpoints or model files required for Malayalam

5)Whether Malayalam support is currently enabled in the released version

Kindly suggest the exact steps to be followed to successfully run the pipeline for Malayalam script recognition.

[anikde]

Hello @dellajoseph1212-bit! Thank you for testing out IndicPhotoOCR.

Malayalam is fully supported within the framework. Since you are not receiving any output predictions, you might be on an older version of the repository where the "auto" script identifier natively defaulted to Hindi and English only, which meant Malayalam bounding boxes were silently discarded or misrouted.

Please git pull the newest version of the repository (or install the latest package). Specifically, regarding your questions:

1) Whether any additional pretrained Malayalam-specific weights need to be downloaded
No manual downloading is required. Under the hood, the pipeline triggers .ensure_model() which automatically downloads both the malayalam.ckpt recognizer and the 12-class script identifier directly on your first execution.

2) If any language configuration parameters must be explicitly set
If you are passing the whole document through .ocr(), simply initialize with identifier_lang="auto"! In the current release, "auto" maps to the 12-class ViT identifier, which covers Malayalam natively.

from IndicPhotoOCR.ocr import OCR
ocr_system = OCR(verbose=True, identifier_lang="auto")
results = ocr_system.ocr("your_malayalam_image.jpg")

Alternatively, if you know the document is purely Malayalam, passing identifier_lang="malayalam" will force the loading of our tri-lingual (Hindi/English/Malayalam) identifier which limits inference permutations purely to those three languages.

3) Required input image format or preprocessing steps
No manual preprocessing is required! Pass arbitrary raw scene images (.jpg / .png) directly into .ocr() and the TextBPN++ backbone will handle text detection and automatic polynomial cropping for the scripts.

4) Any additional checkpoints or model files required for Malayalam
None. Both the detection framework, the script identification backbone (hindienglishmalayalam.zip / 12_classes.zip), and the recognition checkpoint (malayalam.ckpt) are resolved internally.

5) Whether Malayalam support is currently enabled in the released version
Yes! Native support is fully enabled across all three core modules (Detector, ViT Identifier, and PARseq Recogniser).

(Pro-tip: If you strictly want to extract the literal text from a pre-cropped word, you can bypass the detector and identifier entirely by directly hooking the recogniser: text = ocr_system.recognise("crop.jpg", "malayalam"))

Please verify that you are on the newest commit and share your console warnings if the error persists!