Learning Deep Latent Representation of Color Images Using Autoencoders for Efficient Image Retrieval

The content-based image retrieval (CBIR) systems exploit the visual information present in the images to find a suitable match similar to the human visual saliency mechanism. The rise of deep learning methods accomplished notable results in learning efficient image descriptors over traditional featu...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Annual IEEE India Conference S. 1 - 6
Hauptverfasser: Kale, Mandar, Mukhopadhyay, Sudipta
Format: Tagungsbericht
Sprache:Englisch
Veröffentlicht: IEEE 19.12.2024
Schlagworte:
Autoencoders
CIFAR-10
CIFAR-100
Color
Deep Autoencoder
Deep learning
Feature extraction
Fuzzy Class membership
Image retrieval
Performance analysis
Representation learning
Retrieval
Training
Visualization
ISSN:2325-9418
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:The content-based image retrieval (CBIR) systems exploit the visual information present in the images to find a suitable match similar to the human visual saliency mechanism. The rise of deep learning methods accomplished notable results in learning efficient image descriptors over traditional features. These methods automatically extract the salient features and reduce the dimension to get the useful representation of raw image data. In this light, the article proposes a feature learning framework using a deep stacked sparse autoencoder (SAE) and convolutional autoencoder (CAE) models. The autoencoder learns the meaningful approximation of the raw image data following unsupervised training. The study shows the impact of unsupervised training of autoencoder networks on the efficacy of the learned latent features using large-size CIFAR-10 and CIFAR-100 databases. Different model architectures are evaluated to find a suitable architecture for efficient feature extraction of natural color images. The proposed autoencoder networks achieve a nearly 90% reduction in size while maintaining accuracy with a simple distance-based approach. Further, the efficacy of the learned features is tested with two classifier-based retrieval methods using a trained softmax classifier. The experimental evaluation shows that the proposed method SAE and CAE model achieves promising improvements and highly competitive retrieval performance with a large-size CIFAR-10 database.
ISSN:2325-9418
DOI:10.1109/INDICON63790.2024.10958253