Autoencoder Based Image Augmentation and Reconstruction Analysis Using a Reward-Penalty Method

Abstract

Deep convolutional autoencoders hold significant promise for image restoration and feature learning in data-constrained domains. However, their efficacy varies across different corruption types. In this study, we systematically evaluate a plain convolutional autoencoder’s ability to reconstruct images subjected to five representative augmentations: speckle noise, salt-and-pepper noise, Gaussian noise, selective inversion, and patch shuffling. We introduce a novel, layer-wise feature-matching reward metric to accurately quantify augmentation difficulty. We trained and tested our model on a diverse dataset of natural scenes, measuring the reconstruction fidelity via the peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and the proposed reward across the network depth. Our results reveal that noise-based distortions are effectively mitigated, achieving PSNR values near 20 dB and SSIM scores of approximately 0.50. In contrast, spatially disruptive corruptions: selective inversion and patch shuffling, yield substantially lower PSNR (≈13 dB) and SSIM (<0.25). The reward metric correlates closely with quantitative performance, providing a coarse but expedient ranking of augmentation difficulty. Qualitative analyses underscore the autoencoder’s proficiency at preserving global structure under sparse corruption . We discuss implications for architectural enhancements (e.g., skip connections, attention modules) and advanced loss formulations (e.g., perceptual or adversarial losses) to bridge these performance gaps. We establish a systematic framework for evaluating augmentation strategies, informing the design of robust, data-efficient models capable of handling diverse real-world distortions.