Portfolio

Neural Network accelerator

License plate recognition

Let’s see how the NightVision network and the Denoising network can be used to optimize the license plate recognition of following image.

Taken from an image with a longer exposure time, we know that the license plate is 5NON243

Car detection

The original image above is a 20MP image. Processing such an image requires a lot of processing power. To recognize the license plate, we first detect the car on a low resolution image with the a fasterrcnn_mobilenet_v3_large_320_fpn network. We see that by applying the NightVision network first, the confidence of car detection increases from 90% to 97%.

If you look closely to the images, you can also see that the bounding box of the car on the original image is a little bit too small, not completely including the license plate.

License plate recognition

Let’s further extend the image processing pipeline. We use following license plate recognizer to identify the license plate on the original and derived images.

Original	Denoised
Night Vision	Denoised + Night Vision

Construction of a good image processing pipeline, when done in the correct order, can improve the results of classification and detection networks.

This particular pipeline contains 4 AI components and some very simple image preprocessing steps.