Master's Thesis : Evaluation and Integration of Deep Learning Architectures for Automatic Defect Recognition

Abstract

Industrial control has take more and more importance as pieces are often designed to meet the technical requirements while being the more lighter and cheaper. In the same time, an operator is responsible for manually inspect and detect defective part from non-defective ones, which can less and less be achieved in reasonable times for large amount of pieces. This feeds the demand for innovative and efficient methods to inspect and discriminate between defective and non-defective parts. With the impulsion of companies like X-RIS, Euresys and Optrion, the Walloon Region has launched the ADRIC project. This project aims at developing numeric solutions to face the aforementioned problem. The goals of this thesis are to investigate deep learning solutions, to implement these solutions and to evaluate these solutions over x-ray images of industrial pieces. Two datasets are provided by the ADRIC team and hosted on Cytomine[19]. The GDXray[21] dataset is also used in the experiments. One evaluates the Mask R-CNN[9] model within the Detectron2[24] framework. One also builds our own custom models and evaluates those models, to assess the influence of several parameters on their accuracy. the best model is then integrate to the Cytomine Research platform. One follows the integration protocol and release an inference algorithm based on a pre-trained model, previously selected.