Cross Synergetic Mobilenet-VGG16 for UML Multiclass Diagrams Classification

Abstract

Unified Modeling Language (UML) diagrams are a standard modeling language to represent design of software systems. Specifically, UML includes several types of diagrams permitting to assist developing and designing efficiently any software. Additionally, class diagrams (i.e., class diagrams, activity diagrams, sequence diagrams, and use case diagrams) are the most widely used UML diagrams as a standard modeling language for object-oriented software. However, manually classifying UML diagrams is time-consuming and requires an important effort. In addition, there is a need to automate UML class diagrams classification in order to assist researchers, software developers, and academicians to efficiently study and analysis software. One solution can be considered is to use DL-based classification methods as these methods have gained special popularity in different computer vision tools. However, while the majority of innovatory deep-learning efforts using Convolutional Neural Networks (CNNs) focus on improving more refined and strong architectures (e.g., Mobilenet, VGG16, ResNet, U-Net, GANs), there is very restricted work on how to concatenate dissimilar CNN architectures to enhance their relational learning of CNN-to-CNN interactions. In this paper, our purpose is to address this limitation in order to automate the UML diagrams classification. For this reason, we propose a Cross Synergetic Mobilenet-VGG16 (CS-Mobilenet-VGG16), which aims to tackle two crucial problems in computer vision classification and ensemble CNN learning: (1) a bi-directional flow of information between two CNNs (Mobilenet-VGG16) where information proceeds in a directional method from Mobilenet to the VGG16, and (2) synergetic variability across UML diagrams. Our Cross Synergetic Mobilenet-VGG16 significantly (p<0.05) outperformed several conventional CNN architectures (Mobilenet and VGG16).