Research Engineer Detection of complex visual anomalies using very weakly supervised learning: Studies, applications and transposition of Visual Large Model methods to the industry of the future.

This recruitment is part of the PostGenAI project led by Sorbonne University. The project is supported by the French National Research Agency (ANR) as part of the government’s France 2030 investment plan.

Missions

• To carry out a rigorous scientific study aimed at comparing the performance of deep learning models in detecting complex visual anomalies.

• Take charge of the entire study, define the evaluation criteria and model deployment configurations, and carry out the associated experiments.

• Contribute to the drafting of scientific articles presenting the study’s findings and offering recommendations on the adaptation and deployment of VLM models within the industry.

ACTIVITIES

The subject addresses the general issue of complex visual anomaly detection in industry. Industry needs to adopt trustworthy inspection tools that can be rapidly deployed in production. The most effective supervised AI solutions rely on a prior, lengthy, and costly image-labelling process by process experts (several hundred labelled images). Complex visual anomalies are: (1) logical defects where models must detect and verify the relative positioning logic between different objects or entities within the scene. Examples include checking for the absence/presence of parts or components, detecting wiring errors, etc.; (2) texture defects, such as opacities, cracks, tear-offs, etc.

• The primary activity of the role will be to study and compare the performance of VLM models in detecting these anomalies. Cost criteria will be taken into account (related to inference, labelling, prompt formulations, power consumption, etc.). Different model types (open-format, proprietary, etc.) must be considered for the study.

• Based on the results obtained, recommendations on transposing these methods to industry will be proposed and then promoted to UTC's industrial partners.

• Writing, submitting, and publishing in scientific journals are expected.

ADDITIONAL information

Application dates

From 21/09/2026 to 20/10/2026

Contract type and expected start date

Fixed-term contract – expected duration of 18 months – to start in October 2026

Gross monthly salary

Depending on experience and funding

Working hours

37 hours and 30 minutes per week – 1,607 hours per year

Scientific background / project summary

PostGenAI@Paris sets out to anticipate the next breakthroughs in artificial intelligence and to take ownership of their scientific, societal and ethical implications — at the very point where the boundary between technology and human intelligence steadily blurs.

https://postgenai.sorbonne-universite.fr

The project focuses on CAP 1.6 AI for Industry, which aims to develop methods for explainability and uncertainty quantification to make AI usable, auditable and truly operational in risk-related industrial decision-making.

https://postgenai.sorbonne-universite.fr/pac/cap-1.6

Although a large number of solutions exist (local methods, global methods (Carvalho et al., 2024)), there is still room for improvement regarding false alarm rates. Initial findings from the benchmark study on AutoVI show that for logical defects (such as wiring), the false alarm rate ranges between 5% and 20%. It is therefore vital to introduce new methods that help explain the reasons behind a false trigger. Consequently, one key bottleneck addressed in this project is reducing false alarm rates and improving the explainability of results for images incorrectly classified as defective.

Skills

• Ability to design and implement a rigorous experimental scientific protocol

• Solid foundation in Python programming and an understanding of AI architectures based on LLMs and VLMs

• Knowledge of visual defect inspection for industry

Diploma

Engineering degree or Master's degree

Research Field

Industrial Engineering, Computer Engineering