A fleet manager reached out to us a few months ago with a problem I hadn't seen framed quite this way before. He had hundreds of scanned pages from Hino Melpha parts catalogs — images, PDFs, photographs of printed sheets — and absolutely no way to use them efficiently. Every time his team needed to cross-reference a part number, someone had to manually flip through files, squint at low-resolution scans, and transcribe numbers by hand into a spreadsheet. Hours of work. Constant errors. Real operational cost.
That's the kind of problem we love at Helion 360. It sits at the intersection of data strategy, process design, and practical tooling. And solving it taught me a lot about how to approach parts catalog digitization for specialist commercial vehicles like the Hino Melpha — a bus model with a fairly specific aftermarket ecosystem and documentation that hasn't always been well-digitized.
Why Hino Melpha Parts Documentation Is Particularly Tricky
The Hino Melpha is a midsize route bus used across Southeast Asia and parts of the Middle East. Its parts documentation tends to exist in a few forms: original printed catalogs, scanned copies of those catalogs, and occasionally partial digital exports from dealer systems. None of these formats talk to each other naturally.
The images themselves present several challenges:
- Variable scan quality — some pages are crisp, others are clearly third or fourth-generation photocopies
- Mixed languages — Japanese part descriptions alongside English and sometimes Bahasa or Arabic
- Diagram-heavy layouts — exploded assembly diagrams with callout numbers that reference a separate text table on the same or adjacent page
- Non-standard table formats — column headers change between catalog versions, and some sections use freeform notation rather than strict rows
Getting clean, usable Excel data out of this requires more than just running an OCR tool and hoping for the best.
The Workflow I Built to Extract Structured Data
Here's the actual process I used to take raw Hino Melpha parts images and produce a clean, sortable, searchable Excel file. This isn't a one-click solution — it's a repeatable workflow that produces reliable output when followed carefully.
Step 1: Image Preprocessing
Before any extraction happens, image quality determines everything downstream. I ran each image through a preprocessing step using a combination of tools — primarily Adobe Acrobat's built-in enhancement for PDFs and a batch processing script using ImageMagick for standalone image files.
The key adjustments were: increasing contrast to sharpen text edges, deskewing pages that had been scanned at slight angles, and despeckling to remove noise artifacts from older prints. For pages that were genuinely too degraded, I flagged them for manual review rather than letting bad data pollute the output.
Step 2: OCR with Layout Awareness
Standard OCR treats a page as a stream of text. That's useless for parts catalogs, where the spatial relationship between a callout number in a diagram and its corresponding row in a table is what gives the data meaning.
I used a layout-aware OCR approach, leveraging Google Document AI for the higher-quality scans and Tesseract with custom table detection configurations for bulk processing. The goal at this stage wasn't perfect text — it was identifying table structures: which text blocks belonged to headers, which to rows, and which were diagram annotations versus catalog data.
Step 3: Field Mapping and Normalization
Once I had raw extracted text with positional data, the next step was mapping it to a consistent schema. For Hino Melpha parts data, the fields I standardized across catalog versions were:
- Part Number (Hino OEM format)
- Part Description (English, translated where needed)
- Assembly Group / Section Reference
- Quantity per Assembly
- Applicability Notes (model year, variant)
- Supersession or Cross-Reference Number (where available)
Normalization was the most time-consuming part. Part numbers in older Melpha catalogs sometimes use spacing and hyphenation inconsistently. I wrote a Python script to enforce a consistent format and flag anomalies for human review.
Step 4: Diagram-to-Table Reconciliation
This is the step most automated tools skip entirely. In a parts catalog, exploded diagrams show how components fit together physically, with callout numbers pointing to specific parts. Those callout numbers correspond to rows in the adjacent table.
I built a reconciliation step that cross-checked extracted table rows against identified callout numbers from the diagram regions. Any callout without a matching table row — or any table row without a corresponding callout — was flagged. This caught about 8% of records that would have otherwise been silently wrong.
Step 5: Export to Excel with Structured Formatting
The final output wasn't just a raw data dump. I structured the Excel workbook to actually be useful for the fleet manager's team:
- Separate sheets per assembly section (engine, transmission, body panels, etc.)
- A master lookup sheet with all parts in a single searchable table
- Conditional formatting to highlight superseded part numbers
- A data validation layer so future manual entries follow the same format
The client went from spending 3-4 hours per parts lookup session to under 15 minutes. That's the actual business outcome this workflow delivers.
Tools Summary
For anyone replicating this process, here's what I actually used:
- Adobe Acrobat Pro — PDF enhancement and initial OCR pass
- Google Document AI — layout-aware extraction for complex table pages
- Tesseract OCR — bulk processing with custom psm configurations
- Python (pandas, openpyxl, regex) — normalization, reconciliation logic, Excel export
- ImageMagick — batch image preprocessing
When This Approach Makes Business Sense
If you're managing a fleet with older Hino Melpha units, or you're a parts distributor dealing with physical or scanned catalog inventory, the ROI on digitization is usually clear within the first month. The bigger question is whether to build this capability internally or bring in a team that already has the workflow developed.
For one-time catalog conversions, it often makes more sense to work with a specialist. For ongoing catalog management — where new parts bulletins come in regularly and the dataset needs to stay current — building an internal process with the right tooling is worth the investment.
Either way, the starting point is the same: treat the image-to-data conversion as a structured engineering problem, not a copy-paste task.