Car Damage Detection on VGG-16

For this we used a CNN model, VGG16[1]. We used our own dataset that we had prepared and trained the VGG16 model ourselves. Our study consisted of 4 main questions:

• Is it a car?
• Is there any damage on the car?
• In which part of the car is the damage?
• What is the level of damage?

We prepared and trained our own datasets within these 4 questions.

Data Sets

Data Sets	Training	Validation
Is it a car?	920	230
Is there any damage on the car?	1840	460
In which any part of the car is the damage?	976	171
What is the level of damage?	979	171

We used our original dataset, which consists of a total of 5,757 photographs.

Is it a car?

Data Sets	Training	Validation
Car	920	230

The result we will get here is only querying whether there is a car or not.

Is there any damage on the car?

Data Sets	Training	Validation
Damaged	920	230
Undamaged	920	230

We have done this training in order to determine whether there is a similar damage at this stage and to continue with the other stages according to the result.

In which any part of the car is the damage?

Data Sets	Training	Validation
Front	418	73
Rear	287	50
Side	271	48

our aim here is to determine which part of the car the damaged area in the photo belongs to. We trained our parameters as front, back and side.

What is the level of damage?

Data Sets	Training	Validation
Minor	278	48
Moderate	315	55
Severe	386	68

Now that we know the damage is and where it is, it's up to us to separate the level of damage we set ourselves.

Accuracy Rating

Training	Accuracy Rating
Is it a car?	%98
Is there any damage on the car?	%90
In which any part of the car is the damage?	%70
What is the level of damage?	%66

Result Exp

License

Authors

• @ayberkgezer
• @TediTae