Boosting Feature Learning in Autoencoders with Convolutional Layers
Volume 20, Issue 1, 2026
Download| Author(s): |
Muhammad Usama Yousaf* Faculty of Information and Communication Technology, Universiti Tunku Abdul Rahman, Malaysia, its.usamayousaf@gmail.com Rabia Khan Department of Management Science, The University of Haripur, Pakistan, rabiabba43@gmail.com Aiman Sohaib Department of Management Science, The University of Haripur, Pakistan, aimansohaib223@gmail.com Ayesha Saddiqua 3Institute of Computing, Kohat University of Science and Technology, Pakistan, kim.ayeshasaddiqua@gmail.com Mahnoor Izzat Faculty of Information and Communication Technology, Universiti Tunku Abdul Rahman, Malaysia, mahnoorizat187@gmail.com Sana Hassan Faculty of Business and Finance, Universiti Tunku Abdul Rahman, Malaysia, sanahassan@1utar.my Muhammad Adnan Kaim Khani Faculty of Information and Communication Technology, Universiti Tunku Abdul Rahman, Malaysia, adnankk12@1utar.my |
|---|---|
| Abstract | This research proposes a Convolutional Autoencoder (CAE) as an alternative to traditional Autoencoders (AEs) for improving both image reconstruction quality and feature representation. While conventional AEs are effective for dimensionality reduction and unsupervised feature learning, they often struggle to accurately reconstruct compressed images, which leads to a loss of spatial information and, in turn, weaker feature representations and reduced classification performance on downstream tasks. To address this limitation, the proposed CAE integrates convolutional layers into its encoder and decoder, enabling the model to capture local spatial hierarchies and translational invariance that are critical for image data. Supplementary compression layers are further introduced to refine the latent representation, allowing for more efficient encoding with minimal information loss. Together, these design choices improve the model’s ability to reconstruct images with higher fidelity while extracting more discriminative features. The proposed CAE is evaluated against three established models, D-CAE, ScatNet, and PCANet, on the MNIST dataset, where it achieves 96% accuracy, outperforming ScatNet (93%), PCANet (92%), and D-CAE (92%). It attains this result with a comparatively lightweight architecture, suggesting a favorable balance between accuracy and computational cost, although a detailed runtime comparison is left for future work. These findings indicate that convolutional enhancements can make autoencoders more effective for image reconstruction and classification tasks that require both accuracy and efficiency; since the evaluation is limited to the MNIST benchmark, further testing on more complex, real-world datasets would be needed to confirm how broadly these benefits generalize. |
| Keywords | aConvolutional Neural Network, Autoencoder, ScatNet, and PCANet |
| Year | 2026 |
| Volume | 20 |
| Issue | 1 |
| Type | Research paper, manuscript, article |
| Journal Name | Journal of Information & Communication Technology | Publisher Name | ILMA University | Jel Classification | - | DOI | - | ISSN no (E, Electronic) | 2075-7239 | ISSN no (P, Print) | 2415-0169 | Country | Pakistan | City | Karachi | Institution Type | University | Journal Type | Open Access | Manuscript Processing | Blind Peer Reviewed | Format | Paper Link | https://jict.ilmauniversity.edu.pk/journal/jict/20.1/4.pdf | Page | 28-34 |