Car Damage Classification By: David Parra Adrian E. Gonzalez Motivation Wanted to work on a project that involved being able to identify and determine the damage in a car of a given an image .
Background Current visual inspection and validation is done by somebody and because of this, it tends to delay the claim process. Car insurances tend to waste resources due to claim leakage claim leakage: lost dollars through claims management inefficiencies
This Claim leakage can be the difference between the actual claim payment made and the amount that should have been paid Problem Attempting to obtain a higher accuracy using CNN in Car Damage Classifications, while using a limited set of Data.
Goals Being able to Compare 2 types of Cars Such as Damage vs Whole Being able to Compare at least 3 levels of damages Such as Minor vs Moderate vs Severe Being able to determine the side of the car that is damaged
Such as Front vs Rear vs Side Related Works Previous works used different methods based off of CNN in attempting to solve the problem, ultimately achieving between 80%90% accuracy using CNN.
https://www.researchgate.net/publication/322671990_Deep_Learning_Based_Car_Damage_Classification Data Data Set in total is more than a thousand images Dimensions are (160, 160, 3) 8 classes Data is separated into this cases for the demo
Data Set 1: Damage, Whole Data Set 2: Front, Rear, Side Data Set 3: Minor, Moderate, Severe Each Data Set has its own Training and Validation Data in which it was split into 80%-20%, where 80% was used for training and 20% was used for testing.
Whole Damaged Minor Damage
Moderate Damage Severe Damage Method(intro) - Training a neural network from scratch requires: - A lot of data
- A lot of time - A lot of processing power - We use Transfer learning from pretrained models Method(cont.)
PRETRAINED CNNs (keras) Xception
VGG16 VGG19
ResNet, ResNetV2, ResNeXt InceptionV3
InceptionResNetV2 MobileNet
MobileNetV2
DenseNet NASNet MODEL
Training Parameters(1) images Training Parameters(2) Optimizers
fine tuning Experimental Results(1) 3 cases where the architecture worked ok Damage vs Whole Minor vs Whole
Moderate vs Whole Experimental Results(2) 3 cases where the architecture didnt work as expected Severe vs Damage
Damage vs Front Moderate vs Minor Experimental Results(3) Front vs Rear vs Side
Minor vs Moderate vs Severe Discussion about Results and Performance -Overall accuracy is between .6 and .8 -Accuracy remains low unless classes are very different -Fine tuning helped in most cases to gain some accuracy
Future Works and Plans Test with more pretrained models Extend the dataset Try different classifiers Try other classes References
K. Patil, M. Kulkarni, S. Karande, "Deep learning based damage classification", 2017 16th IEEE International Conference on Machine Learning and Applications (ICMLA), pp. 50-54, 2017.
