@article{4806,
  author = {Martin Lopez Nores},
  title = {Quantitative Evaluation of Autoencoder Reconstruction Performance on the CIFAR-10 Dataset},
  journal = {Journal of Multimedia Processing and Technologies},
  year = {2026},
  volume = {17},
  number = {3},
  doi = {https://doi.org/10.6025/jmpt/2026/17/3/152-176},
  url = {https://www.dline.info/jmpt/fulltext/v17n3/jmptv17n3_3.pdf},
  abstract = {Autoencoders are foundational unsupervised neural networks that learn compact data representations by
minimizing reconstruction error. This study presents a comprehensive quantitative and qualitative evaluation
of a convolutional autoencoder (CAE) for image reconstruction using the CIFAR-10 dataset. The model,
comprising approximately 1.05 million trainable parameters, compresses 32×32×3 RGB images into a 64-
dimensional latent space before reconstructing them. A multidimensional evaluation framework integrates
pixel-level metrics (MAE, MSE, RMSE), signal quality (PSNR), and perceptual similarity (SSIM), supplemented
by diagnostic visualizations including side by side comparisons, error heatmaps, per-image error
distributions, and class wise performance breakdowns. The autoencoder achieved a PSNR of 24.88 dB and
SSIM of 0.786, indicating acceptable signal quality and good structural preservation. Statistical validation
using one-way ANOVA (F(9,9990) = 17.84, p < 0.001) and Tukey HSD post hoc tests revealed significant
class-dependent reconstruction differences, with vehicle categories (automobile, truck, ship) exhibiting lower
error than animal classes (cat, dog, bird), reflecting the influence of semantic complexity on compression
efficacy. Crucially, latent space analysis through PCA, t-SNE, and UMAP visualizations, alongside clustering
metrics (Silhouette Score = 0.74, Calinski Harabasz Index = 4256.9), demonstrated that semantically
meaningful class organization emerges without explicit supervision, with confusion heatmaps showing 89-
95% diagonal alignment with ground truth labels. This study contributes a robust, reproducible benchmarking
framework for autoencoder evaluation, establishes the relationship between reconstruction fidelity and
latent space semantic structure, and positions the CAE as an effective feature extractor for downstream
tasks such as unsupervised classification, anomaly detection, and transfer learning in computer vision
pipelines.},
}