21 October 2026

According to the developers from the Data Science AI team, Indra OCR recognizes complex documents with dense layouts without falling behind heavier computer vision pipelines. The model contains 1.25 million parameters and occupies a 30 MB file. According to the claimed data, on a modern multi-core processor it processes 49 pages per second. The model runs locally, without dedicated GPU video memory (0 MB).

The test used three tasks with real documents: a scientific paper describing the attention mechanism, an article with an architectural diagram, and an image with a benchmark results table. All computations were performed on the CPU.

Processing Pipeline

Document processing goes through four stages:

1. Orientation detection. The model determines the page orientation and corrects image skew of up to 45°. Alignment is performed before text extraction begins.
2. Layout segmentation. The model separates headings, paragraphs, and tables and calculates precise bounding boxes for each zone.
3. Script detection. The model dynamically determines the writing system of each line. According to the developers, it recognizes Latin, Arabic, and other world scripts.
4. Output generation. The recognized segments are assembled in natural reading order. The output consists of JSON files and tables in Markdown format.

Installation

The model is installed through the official Hugging Face library using the pip install command. The installation was performed in a Google Colab environment. After installation, the help command confirmed that all options of the tool were available.

Launch and Configuration

The test project consists of a main Python file and a configuration file. The configuration specifies:

• names of input files;
• names of output files and required formats;
• number of CPU threads that the model will use.

The main file loads the model from the Indra OCR package and passes it to three functions, each performing a separate task. The functions use the convert module to process input images.

With a value of 4 in the configuration, the model runs four parallel pipelines that perform tasks simultaneously. To reduce processor load, the number of threads is reduced in the configuration file.

The project is launched with the command python main.py. The model also supports operation from the command line. The command with a single file specified runs the full pipeline and outputs a structured report in Markdown. Additionally, a folder for results and the output file format can be specified. The command to launch the local web interface of the OCR viewer opens the recognition visualization in the browser.

Testing

Task 1: SVG Visualization of the Article "Attention Is All You Need"

The model processed the article "Attention Is All You Need" and built a visual SVG file. As a result, the following were extracted:

• mathematical formulas;
• diagram captions;
• tables with color highlighting for different complexity categories.

The elements are arranged in the same order as in the original article.

Processing the 15-page document took 0.2 seconds. The model also generated a JSON report. The report specifies the device (CPU), the number of parallel threads (4), and the absence of video memory usage. The JSON contains structured data on formulas, figures, paragraphs, and table rows.

Task 2: Structured Markdown Report

In the second task, the model processed the same article and performed the following:

• determined the page orientation: 0°, since the document is a scientific article;
• identified all text blocks, including the abstract paragraphs;
• preserved the reading order from top to bottom and left to right;
• output formulas in LaTeX format;
• output tables as Markdown table rows.

Headers and page footers were automatically identified and separated from the main text. The table structure was preserved, including column headers, mathematical formulas, and row boundaries.

Task 3: Benchmark Table Extraction

In the third task, the model extracted the benchmark results table and saved it in JSON. The data is divided into dictionaries for each model: Indra OCR, PaddleOCR, and Tesseract. Separate keys are defined for each dictionary.

The second table contains transformer architectures, along with their training costs and achieved accuracy.

According to the data in the table, Indra OCR works 23 times faster than Tesseract and 15 times faster than PaddleOCR. The model occupies 38 MB and does not use GPU video memory.

Benchmark Results

According to the developers' official data:

According to the developers, on the OmniPage Match benchmark the model scored 71.4%, which is higher than heavier alternatives on mathematical expression recognition.

The math flag adds equation recognition and keeps latency below 400 milliseconds.

According to the developers, PaddleOCR and Tesseract took roughly 41 and 5 seconds, respectively.

Additional characteristics:

• Coordinates: the model returns exact pixel coordinates for each line of text.
• Languages: more than 80 writing systems are supported, including multiple languages on a single page.
• Digital PDFs: text embedded in a PDF is extracted in 1 millisecond. Neural network processing is applied only to embedded scanned images.

Limitations

• Tables without borders. In complex tables without borders and with multi-line cells, the model may incorrectly align row coordinates. The model is better suited for tables with clear borders.
• Handwritten text. When recognizing cursive handwritten text, accuracy drops to 25%. Low-quality images and handwritten texts require specialized computer vision models.

Conclusions

The main advantages of Indra OCR based on the test results:

• operation entirely on CPU without GPU;
• output of bounding boxes and reading order, which eliminates the need for additional post-processing.

According to the test results, the model is suitable for invoices, receipts, and structured business documents with clear tables. For such documents, the model combines high speed, accuracy, and low computational cost.