Workflow for training and serving DL models for image classification and object detection - application to fault detection on electric poles

This technical analysis done at Elvia presents the development of a workflow that allows for training, testing and serving deep learning models that can use any image coming from the image repository and can enrich these images with metadata derived from the results provided by the deep learning (DL) vision models. Using this workflow, an electric pole image classification model and a missing top cap object detection model were trained and served on millions on helicopter inspection images. In addition to obtaining very accurate models, we observe that such a modular workflow have reduced both the time from idea to prototype and the time from prototype to product.