avodovnik / cognitiveservices-vehicles

A Cognitive Services Application for managing visitors in vehicle lanes

cognitiveservices-vehicles

A Cognitive Services Application for managing visitors in vehicle lanes

Azure Functions

The code for Azure Functions used in this system is in the "Functions" directory.

We use a .NET Azure Function (a custom-written web service) to run detection and OCR on number plates, in a process called NPR (Number Plate Recognition).

In the 'Functions/NPR' directory, OCRResponse.cs is merely a helper object generated at json2csharp, for parsing API responses in NPRFunction.cs - in which we call to 2 Microsoft Cognitive Services: Custom Vision (specifically an object detection model), followed by the OCR endpoint of Computer Vision.

The NPR process runs in 3 steps:

1. Given an input image of the front of a vehicle, detect the number/license plate with Custom Vision (code in function DetectPlate() of NPRFunction.cs).
2. Crop the input image to only the number plate using the boundings provided by the detection model (code in function Crop() of NPRFunction.cs).
3. Run this cropped image through the OCR service to recognize characters, then concatenate and return them as a single string (code in function ReadPlate() of NPRFunction.cs).

The results would currently benefit from some post-processing to rid them of confusion resulting from, for example, the region codes that typically appear on plates. In our testing, however, they have returned reasonably accurate estimations of the plate's content, with less than optimal (internet-sourced) input data.