A hybrid deep compression model for enhanced classification of high-dimensional image data

Abstract

High-dimensional image data often contain redundant pixel-level information, leading to increased computational complexity and reduced learning efficiency in machine learning models. To address this challenge, we propose a Hybrid Compressive Feature Reduction Model (HCFRM) that achieves an effective trade-off between computational efficiency and classification accuracy on challenging large-scale datasets such as CIFAR-10, using CSV-shaped image representations. The proposed HCFRM framework employs an optimized configuration consisting of 100 PCA components, a 64-dimensional Autoencoder latent space, and the top 32 Mutual Information-selected features to achieve an effective balance between dimensionality reduction and classification performance. This optimized HCFRM configuration was adopted throughout all experiments to ensure consistent and reproducible evaluation across the benchmark datasets. This adaptive tuning ensures robust performance on flattened high-dimensional representations, demonstrated on CIFAR-10, medical imaging, and hyperspectral datasets, highlighting the model’s flexibility and applicability to various image domains. The decision-making process is guided by the analysis of the expected values of feature vectors projected into a reduced coordinate space through a hybrid sensor processing pipeline. Experimental results demonstrate that HCFRM achieves a 74% reduction in feature dimensionality while providing approximately 28–38% faster training compared with PCA+Autoencoder and Autoencoder-only baselines. Despite aggressive parameter-space compaction, the proposed model attains a classification accuracy of 74%, indicating effective class separability under reduced computational cost. These findings confirm that HCFRM is an efficient and scalable solution for learning large-scale, high-dimensional image data, making it well suited for memory-efficient and high-performance visual learning systems. The proposed framework further enables efficient processing of high-dimensional visual representations for 3D scene reconstruction and volumetric modeling, supporting enhanced computer vision performance. The modular and compressive design of HCFRM makes it suitable for scalable and cluster-based visual learning systems.

Keywords

Hybrid Compressive Feature Reduction Model (HCFRM) / Dimensionality reduction / CIFAR-10 CSV dataset / Feature compression / Principal Component Analysis (PCA)

References

[1] Kiraz H. Deep feature extraction, dimensionality reduction, and classification of medical images using combined deep learning architectures, autoencoder and multiple machine learning models. Elektrik, 2023, 31: 1113-1128

[2] Han S, Mo B, Zhao J, Xu J, Sun S, Jin B. High-quality image compression algorithm design based on unsupervised learning. Sensors, 2024, 24: 6503.

[3] Ali, Wassif K, Bayomi H. Dimensionality reduction for images of IoT using machine learning. Sci Rep, 2024, 14: 7205.

[4] Kurz W, Wang K, Bektas F, et al. Dimensionality reduction in hyperspectral imaging using standard deviation-based band selection for efficient classification. Sci Rep, 2025, 15: 34478.

[5] Mienye D, Swart TG. Deep autoencoder neural networks: A comprehensive review and new perspectives. Arch Comput Methods Eng, 2025, 32: 3981-4000.

[6] Laakom F, Raitoharju J, Iosifidis A, Gabbouj M (2022) Reducing redundancy in the bottleneck representation of the autoencoders. arXiv:2202.04629

[7] Abbasi F. Dimensionality reduction with deep learning classification: A study in an IoT environment. Expert Syst Appl, 2025, 267: 126149.

[8] Andriani S. CNN-AutoMIC: Combining convolutional neural network and autoencoder to learn non-linear features of bytecode images for malware classification. J Syst Archit Evol, 2025, 156: 104507

[9] JayeprokashConvolutional autoencoders for data compression and anomaly detection on small-satellite image payloads. Information, 2025, 16: 690.

[10] Kavitha D, Shobana S, Rajkumar S, Reddy GR, Mukherjee P. Hyperspectral image classification using dimensionality reduction deep networks. Int J Intell Syst Appl Eng, 2023, 12: 81-86

[11] YardımcıOptimizing deep learning architectures for SEM image classification through robust integration of NMF and SVD techniques. J Res Eng Sci Manage, 2025, 11: 2

[12] Fei X. Lightweight-VGG: A fast deep learning architecture based on dimensionality reduction and nonlinear enhancement for hyperspectral image classification. Remote Sens, 2024, 16: 259.

[13] Ali, Zhang H, Guo Q, et al. Dimensionality reduction for images of IoT using machine learning. Sci Rep, 2024, 14: 57385.

[14] Liu Z, Zhang F, Cheng L, Deng H, Yang X, Zhang Z, Zhou C (2023) Simple but effective unsupervised classification for specified domain images: a case study on fungi images. arXiv:2311.08995

[15] Berahmand K. Autoencoders and their applications in machine learning. Artif Intell Rev, 2024, 57: 28.

[16] Archana R, Eliahim Jeevaraj PS. Deep learning models for digital image processing: A review. Artif Intell Rev, 2024, 57: 11.

[17] Ouyang, He S, Zhang G, Luo M, Guo H, Zhan J, Huang Z (2023) Efficient multi-scale attention module with cross-spatial learning. arXiv:2305.13563

[18] Biotto W. Inductive self-supervised dimensionality reduction for content-based image retrieval. Proc SCITEPRESS, 2025, 383: 391

[19] Barsi, Nayak SC, Parida S, Shukla RM. A deep learning-based compression and classification technique for whole slide histopathology images. Int J Intell Syst Appl, 2024, 16: 4517-4526

[20] Yu Y, Wang D, Chen Y, Wang H, Huang M. Contribution Tracking Feature Selection (CTFS) based on the fusion of sparse autoencoder and mutual information. Computers Mater Continua, 2024, 81: 3761-3780.

[21] Roy T, Das P, Jagirdar R, Shhabat M, Abdullah MS, Kashem A, Rahman R. Prediction of mechanical properties of eco-friendly concrete using machine learning algorithms and partial dependence plot analysis. Smart Constr Sustainable Cities, 2025, 3: 2.

Rights & permissions

The Author(s)