Image Difference Quantification Using Autoencoder-Based Latent Representations

Traditional image similarity metrics like Mean Squared Error (MSE) and Structural Similarity Index Measure (SSIM) rely on pixel-level comparisons, which often fail to capture high-level semantic differences or align with human visual perception. This study investigates whether a convolutional autoencoder (CAE) can learn compact, semantically rich latent representations that provide a more robust and perceptually accurate way to quantify image differences.

The researchers developed a convolutional autoencoder architecture consisting of an encoder, a latent representation layer, and a decoder. The model was trained in an unsupervised manner to reconstruct input images, forcing the encoder to compress images into a low-dimensional latent vector that captures essential structural and semantic information. To quantify the difference between two images, the authors compute the cosine similarity between their respective latent vectors. They evaluated this framework across multiple datasets, including dog-cat images, geometric shapes, and outline-shaded image pairs, and validated the results against human Mean Opinion Scores (MOS) from the TID2013 dataset.

The study found that the latent space effectively organizes images into semantically distinct clusters, showing strong inter-class separability and intra-class consistency. For instance, in the dog-cat dataset, 98.4% of inter-class image pairs exhibited similarity scores below 0.5. Furthermore, the latent-space distance showed a positive correlation with human perceptual judgments, confirming that the learned representations are sensitive to perceptually relevant distortions. The method is also computationally efficient, requiring only a single forward pass through the encoder, making it suitable for real-time applications.

This research offers a scalable, semantically grounded alternative to conventional pixel-based metrics. By moving beyond simple pixel intensities, this approach provides a more nuanced way to assess image similarity in tasks such as content-based retrieval, anomaly detection, and automated quality assessment. It demonstrates that deep representation learning can bridge the gap between machine-calculated metrics and human visual perception.