Number plate blur ML (Proof Of Concept)

Automatically detects and blurs the number plate of the car in the picture.

In addition, you can delete the location and time information of the photo.

You can save the processed picture separately from the original file.

You can also transfer the processing results to other app

For example, you can post the picture blurred the number plate with this app to Twitter directly.

Detection of number plate has implemented by machine learning.

All inferences at the time of detection are done on the device. No data is sent to the server on the network.

■How to use

There are two ways to open photos:

・Launch this app, select an image and open it.

・Transfer to this app via the [Share] function of other apps.

…Transfer and blur the photos displayed in the gallery app such as Google Photos with [Share].

…Long tap the image displayed in the browser → transfer with [Share image] to blur.

There are two ways to use the processed photos:

・Save as new file.

…To "/chaenomeles/" folder.

・Transfer to other apps by [Share] of this app.

You can use as follows: Transfer to Twitter, and tweet it. Transfer to Google Maps app, and post it. etc.

■Attention

This app was created to verify if it is technically possible.

Therefore, the quality of the app is not high.

However, I think that it is a good result among similar apps about quality.

Please pay attention to the following points to use.

・This app is often terminated due to insufficient memory.

…It may be difficult to operate with device which have fewer RAM.

…For storage memory instead of RAM, this app will not waste it so do not worry.

・Device may restart by heavy processing.

…When the number plate area is wide, processing becomes extremely heavy, and such symptoms may occur.

…The reasoning itself is not heavy, but since implementation of the blurring processing is not carried out politely, so there is a problem.

・The number plate may not be correctly detected, or it may be detected area which is not a number plate.

Moreover, it can not correct it.

…Since I mainly focus on verifying inference operations, it have not implemented the function to edit results.

・Can not zoom the result image.

There is a button which can activate the viewer, please browse using this.

・App size is too large.

I made it without consideration of size. Excuse me.

■Remarks

◇The data used for inferencing has been learned using 120 photos of cars / K-cars in Japan.

Since the learning result is frozen, there is no effect that the accuracy increases the more to use.

◇Because it has learned only in Japanese number plate, it may be weak for foreign license plate.

However, it seems that learning has proceeded so that it covers "square plate which had written letters fixed to the car", and it will detect even things in foreign countries one.

◇It is weak for trucks and motorcycles, as I made only car / K-cars images as learning dataset.

If you would like to try out various things, you can enjoy using the image by using the browser to search the car test drive article.

If you long tap the image and send it to this app with [Share], you can check the processing result with affordable size images.