Image recognition Image matching: Previously handled damages

Mining Damage Institute Groningen

Automated recognition of previously assessed damages to prevent previously processed damages from being processed again.

Last change on 30th of July 2024, at 15:08 (CET) | Publication Standard 1.0

• Publication category: Impactful algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

• Law
• Nature and Environment

Begin date

11-2023

Contact information

[e-mail] en [e-mail]

Link to publication website

https://www.schadedoormijnbouw.nl/

Responsible use

Goal and impact

A loss adjuster must perform a manual check to determine whether a loss has been treated previously. An expert does this by comparing photos from a loss report with photos from old reports. Assessing whether a loss has been treated previously is a time-consuming and error-prone process. The Groningen Mining Damage Institute (hereinafter: IMG) has therefore developed an algorithm to support the expert in this process.

Purpose: The algorithm supports a loss adjuster in determining whether a loss has previously been assessed by the IMG or its predecessors, such as the Nederlandse Aardolie Maatschappij or the Centrum Veilig Wonen. According to article 2 paragraph 4 of the Temporary Groningen Act, the Institute is not legally authorised to assess a loss if it has previously been dealt with.

Impact: The algorithm tells an expert whether the newly recorded damage matches damage already known to the Institute. The algorithm does not combine personal characteristics, aspects and/or circumstances of you as a person to arrive at an assessment. Only damage that is on the photos is compared with each other.

Considerations

Assessing whether a claim has been previously treated is a time-consuming and error-prone process. Claims experts often have to go through PDFs of more than 100 pages of damage reports to check whether a claim has been previously treated. The algorithm speeds up this process considerably and thus improves the claims handling.

Human intervention

Based on an entered photo of damage, the algorithm produces a list of the ten most similar photos. The algorithm does not select the most similar photo in any case. The algorithm only supports the expert in selecting the photos to be assessed. The expert himself decides whether there is a previously treated damage.

Risk management

At the moment a DPIA (a privacy risk analysis) is being performed on this algorithm. As soon as the risks are known, they will be included in this register. The measures that reduce the risks of using this algorithm will also be included in this register.

Legal basis

Article 2 paragraph 4 Temporary Groningen Act:

The Institute is not authorised to process a claim for compensation for damage if it concerns damage for which:

• a. A damage report or claim has been submitted to the Centrum Veilig Wonen or the operator before 12:00 on 31 March 2017;
• b. A settlement agreement has been concluded by the operator with the injured party or his representative;
• c. The injured party or his/her representative negotiates with the operator with the aim of obtaining compensation for the damage;
• d. A claim is brought before the civil court, unless the claim before the civil court is withdrawn by the applicant with the consent of the defendant; or
• e. The civil court has ruled on the entitlement to and the extent of compensation for the damage.

Links to legal bases

Tijdelijke wet Groningen: https://wetten.overheid.nl/BWBR0043252/2023-07-01

Elaboration on impact assessments

A Data Protection Impact Analysis (a privacy risk analysis) is performed.

Operations

Data

Damage reports (particularly photographs of damage to homes)

Technical design

The algorithm was developed by the IMG. This was done on a Microsoft Azure infrastructure. Python open source packages (including Tensorflow and PyTorch) were used to shape the self-learning capability. Deep learning models were used for the model. A single test dataset was used during development, but the model learns from the feedback of damage experts in production.

External provider

The algorithm was developed in-house.