Car insurance claim cost evaluation and damage detection

Car insurance claim cost evaluation and damage detection


Eurapco is an alliance of large independent European mutual insurance companies, consisting of eight partners operating in 16 countries across Europe. Over 43 million EU citizens trust the Eurapco Partners. Cost evaluation for car claims is a very demanding and expensive process.

Highly skilled experts process the images from car claims with visual analysis and prepare a summary report. It takes a long time to evaluate the claim and often the estimated costs are not accurate enough.

    • Car damage localization (semantic segmentation)
    • Repair cost estimation (regression)
    • Car part detection – define car parts in a photo (multi-label classification)
    • Working with input photos of any size
    •  High level semantic information extraction
    • Dealing with objects at multiple scales (close-ups and whole cars on photos)
    •  Model should be lightweight (GCP with 1 GPU available due to budget limitation)
    • Tensorflow model should be wrapped into TF-estimator API and ready to be deployed to ml-engine.
  1. ok_iconCreated with Sketch.
    Build a prototype using Deep Learning that estimates repair costs for car claims accurately
  2. ok_iconCreated with Sketch.
    The prototype will save time in the cost evaluation of car claims
  3. ok_iconCreated with Sketch.
    GPU acceleration helped to scale the process by greatly increasing the performance of deep learning models

The goal was to build a system using Deep Learning that will estimate repair costs for car claims accurately. This will eliminate errors in manual evaluations by the experts who have no access to the actual data with repair costs or whose knowledge is outdated.


Data preprocessing

Normalization of each image pixel values to be in range [0, 1]

Validation technique

Stratified KFold basing on damage area

Train-time augmentations

  • Random crops
  • Random horizontal flips
  • Sharpen image
  • Gaussian blur
  • Converting images to HSV, increasing each pixels H-value by 10 to 50
Model Architecture

To solve semantic segmentation tasks the UNet based architecture was used with an InceptionResNetV2 encoder.


  • Combination of low-level semantic features with high resolution and high-level semantic features with low resolution for resolution restoring (to have detailed object borders).
  • U-Net was originally used for medical datasets, which are usually small
  • Limited computational resources


UNet model architecture
UNet model architecture

Loss function

CE loss

Takes into account the closeness of a prediction, strongly penalizes the classifier for being 100% sure, but does not capture the characteristics of the topology

Dice loss

Suitable for unbalanced data (harmonic mean of precision and recall)


Applying the Deep Learning approach helps to save time in the cost evaluation of car claims and making more accurate cost estimates.

GPU acceleration helps to scale this process by greatly increasing the performance of deep learning models.

Next story

Seamless launch of 31 Mobile Apps with Manual QA

Read more
What's your challenge? Let us deliver the talent and expertise to help you solve it.
Upload file

(File requirements: pdf, doc, docx, rtf, ppt, pptx)

By submitting completed “Contact Us” form, your personal data will be processed by Ciklum Group and its subsidiary entities worldwide. Please read our Privacy Notice for more information. If you have any questions regarding your rights or would subsequently decide to withdraw your consent, please send your request to us.

Join our team
Do you have a passion for engineering? We’ve been looking for you.
Check open vacancies