Introduzindo MLOps: reconhecimento de placas de licença de caminhões

Abstract

Most Machine Learning (ML) projects do not reach production. Several challenges are faced when managing data and models, so there are no universal solutions. The new paradigm of Machine Learning Operations (MLOps) was created precisely to address these challenges and make the lifecycle of the development of machine learning models more efficient and effective. Therefore, in this work, the main concepts of MLOps were introduced through the presentation of an Optical Character Recognition (OCR) solution with object detection models for the Automatic License Plate Recognition (ALPR) task already validated in the industry. Some MLOps tools were also introduced to exemplify the main concepts of MLOps in practice. The results showed the ease of creating and maintaining data pipelines and making it clear how flexible such processes are so that the data flow is replicable, with minimal changes, in similar contexts, such as OCR of wagon or container codes. Finally, it was detailed which activities can be added to the workflow so that the operations of developing and deploying ML models are more complete.