November, 2022
Arihant Sales has been providing ID Card solutions for over a decade to customers across India. They process and print more than 100K ID cards annually for reputed schools, colleges, and large corporations. They focus on delivering high quality with accuracy and reduced wastage.
Previously, to print the ID cards, the raw images were manually converted and standardized according to predefined ID templates. Each of the raw, straight-out-of-the-camera image was color corrected, checked for brightness, and cropped to a particular size – through a manually intensive process.
We had been consistently achieving high accuracy in our editing process. However, achieving accuracy meant time tradeoff – the process took long hours to complete, which exerted the staff, particularly, more during high demand. This manual job limited our operations and slowed down our scale
Sanjay Jain, CEO, Arihant Sales
They needed a system to automate this manual effort, reduce time to delivery, standardize results, and reduce any error due to manual intervention.
To digitize their editing process, especially the batching, the company built a local server-based application with a Deep Learning model in partnership with MediaAgility, a Google Cloud Partner. The Deep Learning model was trained on Cloud with Machine Learning Engine and was brought offline to run on the application.
Now the local server-based application fetches the raw images from on-prem storage and runs them through the Deep learning model to detect faces. To ensure that the face fits into the ID card’s frame, the model crops the images based on a few predefined image dimensions and cropping ratio. Also, the model detects facial landmarks to auto rotate an image if the face is tilted. Once the cropped images are color corrected and checked for brightness, they are saved to the on-prem storage.
Other prominent features of the ML modeled application are –
- The staff can interject the processing during any part of the editing process
- Same viewport displays previews of the original and edited images for easy comparison
- Each edit iteration presents a revised image without having to reload the page
- The user can trigger events with keyboard support
- For bulk processing, source and destination directories can be set. The user can create presets based on a number of configurable threshold values that can be used to customize the output.
Image loaded into the application
Image processed – cropped, and color and brightness corrected
Our business was not Cloud enabled but MediaAgility (now acquired by Persistent) met us where we were with their agile and focused approach. They gave us the best of both worlds – we leveraged Cloud for heavy compute loads like training a Deep Learning model, and ran the application on our on-prem environment. MediaAgility engineers integrated the two environments to ensure a steady process flow. And, Google Cloud Platform gave us the infrastructure scale required to train the Deep Learning model. Earlier, manually processing each image took a maximum of 2 minutes. But now we do that in under 10 seconds. That has led to nearly an 80% decrease in the editing time
Sanjay Jain, CEO, Arihant Sales
Arihant Sales has achieved process automation, reduced human errors, amped up the speed of execution, standardized image crops, and reduced wastage with the ML modeled application.
For the next phase, Arihant Sales is working with us to perform Optical Character Recognition (OCR) on handwritten digits and automate its manual efforts of converting hardcopies of forms into digital copies.
Technology Used
- GCP
- Vertex AI
- BigQuery
