Search for a command to run...
Periodic inspection of engineering structures is essential in order to assess their condition, carry out maintenance work, and ensure user safety. The first step in the inspection process is to list all the defects present on the structure. This essential task can sometimes be very time-consuming. To facilitate this survey, the use of automatic image-based defect detection methods is particularly useful.Traditional supervised methods are not entirely satisfactory because they require a large number of annotated examples. Indeed, detecting anomalies on civil engineering structures is a complex task due to the wide variety of defect types and their low occurrence in the available data, which may come from recent structures with few problems. New methods capable of learning under these various constraints must therefore be developed as an alternative to supervised methods.Our study therefore focuses on so-called ``unsupervised'' methods, which only require images without disorder to perform their learning. They aim to model these healthy images and then detect elements that deviate from this normality as defects. After studying the different learning paradigms underlying these methods, an evaluation comparing these different approaches is carried out on a database of civil engineering structures images, acquired during the thesis. Models based on Normalizing Flows, recognized for their performance and robustness, are then specifically explored in order to optimize their detection capabilities. Several investigations are conducted to propose a model better suited to images from engineering structures. First, an ablation study focusing on the essential components of the Normalizing Flow model is conducted, both on reference datasets for anomaly detection and on databases of images of structures, to determine the most effective structure. Next, pre-training based on a pretext task using synthetic anomalies generated by a process built on Poisson interpolation is proposed to adapt the model's feature extractor. The impact of this pre-training varies depending on the dataset. Finally, a Normalizing Flows architecture integrating contrastive learning from synthetic defects is constructed to enhance the model's discriminative power. This approach improves the performance of the Normalizing Flows model on most datasets. This latter structure will ultimately be selected as the anomaly detection model for engineering structure images.