This CIFRE-funded PhD project makes it possible to bring to fruition a collaboration between the Imagine team at LIRIS, specializing in visual data processing and analysis, and the Veolia group, specializing in ecological transformation, waste recovery, and wastewater conveyance.

An increasing number of applications require the acquisition of large quantities of images that must be analyzed in real time in order to trigger actions. Automated processing of these large volumes of data can benefit from recent advances in deep learning. To ensure optimal and safe management of wastewater conveyance infrastructure, CCTV inspection—using a trolley equipped with a camera inserted into a pipe—can be used. The process is mostly manual and relies on the operator’s expertise: they view the images live, detect and identify observable defects (cracks, breaks, defects at joints between pipe sections, etc.), and produce an inspection report. Similarly, in the context of solid-waste collection and sorting, quality control of complex waste streams is carried out using images in which experts identify intrusive objects.

Improving these quality-control processes and anomaly identification in images and video is therefore a crucial challenge. Various deep-learning, statistical, and image-processing techniques have been implemented to meet this need. However, these techniques require large volumes of annotated images, and the class imbalance between the objects or defects to be identified represents a real challenge for operational deployment.

The aim of the PhD project is to explore different approaches (image synthesis, hierarchical annotation, self-supervised learning, generative adversarial networks, etc.) to generate representative annotated images using information transferred from other contexts or weak supervision.