Code generation system based on MDA and convolutional neural networks.

Abstract

Introduction: The software industry has rapidly evolved with high performance. This is owing to the implementation of good programming practices and architectures that make it scalable and adaptable. Therefore, a strong incentive is required to develop the processes that initiate this project.

Method: We aimed to provide a platform that streamlines the development process and connects planning, structuring, and development. Specifically, we developed a system that employs computer vision, deep learning, and MDA to generate source code from the diagrams describing the system and the respective study cases, thereby providing solutions to the proposed problems.

Results and discussion: The results demonstrate the effectiveness of employing computer vision and deep learning techniques to process images and extract relevant information. The infrastructure is designed based on a modular approach employing Celery and Redis, enabling the system to manage asynchronous tasks efficiently. The implementation of image recognition, text analysis, and neural network construction yields promising outcomes in generating source code from diagrams. Despite some challenges related to hardware limitations during the training of the neural network, the system successfully interprets the diagrams and produces artifacts using the MDA approach. Plugins and DSLs enhance flexibility by supporting various programming languages and automating code deployment on platforms such as GitHub and Heroku.