MANAS-DRISHTI OCR

MANAS-DRISHTI OCR is a compact optical character recognition model that reads text in images — including engineering and technical annotations (dimensions, tolerances, GD&T symbols like Ø ° ±).

Author: @k_k_kaundal on X

What it does

• 🔍 Reads text from images (OCR)
• 📏 Handles engineering & technical text: dimensions, tolerances, fits, GD&T symbols
• ⚡ Small and fast — runs on CPU

File

• drishti-ocr.pt — the model weights (PyTorch checkpoint)

Usage

Load the checkpoint and read text from an image:

import torch
from huggingface_hub import hf_hub_download

# 1) download the weights
path = hf_hub_download("kaundalvip/manas-drishti-ocr", "drishti-ocr.pt")

# 2) load the checkpoint
ckpt = torch.load(path, map_location="cpu")
# ckpt["state"] -> the OCR model weights
# ckpt["img_h"] -> the input image height the model expects

# 3) build the OCR reader from the weights, prepare your image as a
#    grayscale tensor of height ckpt["img_h"], and run it to decode text.

The checkpoint contains the model weights (state) and the expected input image height (img_h). Feed a grayscale image crop and the model returns the recognized text.

Performance

• Character error rate (CER): ~0.001 on held-out text
• Exact-match: ~90–96%

Example

Image of text: "Ø25 H7 ±0.05"
Model reads:   "Ø25 H7 ±0.05"

Connect

Built by K K Kaundal — follow on X: @k_k_kaundal