EnsemV3X: a novel ensembled deep learning architecture for multi-label scene classification.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Convolutional Neural Networks (CNNs) are pivotal for image classification, often utilizing pretraining on large datasets like ImageNet.
• ImageNet, a vast dataset with 15 million images across 22,000 categories, facilitates effective transfer learning.
• Transfer learning with pretrained models enables efficient and accurate development of computer vision models.

Purpose of the Study:

• To demonstrate multilabel scene classification using established CNN architectures.
• To compare the performance of VGG16, VGG19, ResNet50, InceptionV3, and Xception for scene recognition.
• To introduce and evaluate a novel ensemble model, EnsemV3X, for enhanced scene classification.

Main Methods:

• Utilized five distinct CNN architectures: VGG16, VGG19, ResNet50, InceptionV3, and Xception.
• Leveraged ImageNet weights available in the Keras library for model pretraining and fine-tuning.
• Developed and tested a proposed ensemble model, EnsemV3X, integrating features from multiple architectures.

Main Results:

• Comprehensive performance comparison of the five evaluated architectures on multilabel scene classification.
• The proposed EnsemV3X model achieved a high accuracy of 91%.
• EnsemV3X demonstrated superior performance compared to state-of-the-art models like InceptionV3 and Xception, despite having fewer parameters.

Conclusions:

• Multilabel scene classification can be effectively achieved using deep learning architectures pretrained on ImageNet.
• The EnsemV3X model represents a significant advancement, offering improved accuracy and efficiency in scene recognition.
• The study highlights the potential of ensemble methods in pushing the boundaries of computer vision applications.