Skip to content

Post-Report PyTorch Autoencoders

Not Part of the Main Project Report

The notebooks and models described in this document (05_unified_perceptual_autoencoder.ipynb, 05b_unified_perceptual_autoencoder_eval.ipynb, and mina_autoencoder_final_improved_and_ssim.ipynb) were added during post-report exploratory phases. They are not part of the official project deliverables or the primary report. The main project model remains the Keras Convolutional Autoencoder (CAE).

1. Context & Motivation

Following the completion of the main project report and the Keras CAE implementation, further exploratory research was conducted using PyTorch to evaluate the impact of advanced structural loss functions and Masked Image Modeling (MIM) on fine-grained industrial components (such as screws).

Subtle technical note regarding the earlier Keras CAE: During the post-report review, it was noticed that the Keras model accidentally utilized a combination of SSIM+L2 loss, rather than a strictly optimized perceptual formulation. This PyTorch pipeline allowed us to explicitly test MS-SSIM combined with L1 and Perceptual Loss (VGG/ResNet features).

2. The Unified Perceptual Autoencoder

The notebook 05_unified_perceptual_autoencoder.ipynb implements a ConvAutoencoder utilizing ELU activations and a narrow bottleneck (64 channels at 16x16 resolution) to act as a strong regularization mechanism.

Key Objectives Investigated

  • MS-SSIM + L1 + Perceptual Loss: Evaluated against standard MSE. MS-SSIM computes structural similarity across multiple scales, ignoring minor misalignments and lighting variations while forcing the model to reconstruct physical component structures accurately.
  • Masked Image Modeling (MIM) Ablation: Explored dynamic patch-masking (MIM) to force the model to "inpaint" missing patches.

Findings on Fine-Grained Defects (Screws)

The ablation study revealed that MIM is detrimental to fine-grained defect detection. For items like screws, if a thread is scratched or malformed, the MIM pipeline interprets the defect as "missing information" and perfectly hallucinates a normal screw thread. Because the reconstruction perfectly corrects the defect, the model yields a zero anomaly error over the defective region, severely reducing the Hit Error Rate (HER).

Therefore, the recommended baseline for these components avoids MIM and strictly utilizes Perceptual Loss with MS-SSIM for shape-aware robustness.

3. Strict Evaluation Enhancements

The evaluation script (05b_unified_perceptual_autoencoder_eval.ipynb) was refined to match the rigorous statistical boundaries established in the dataset benchmarks:

  • Image-Level Thresholding: Thresholds for binary anomaly classification are now correctly calculated using the 99th percentile of image-level max pixel errors in the normal validation set, replacing the noisy all-pixel percentile.
  • AUPIMO Integration: Formal integration of anomalib.metrics.AUPIMO with standardized False Positive Rate bounds ([1e-5, 1e-4]) and 50,000 threshold bins.
  • Hit Error Rate (HER): Computation of the Top-1% Defect Capture rate to measure true spatial localization hit rates.

4. Mina's Improved Notebook

Also merged into this branch is mina_autoencoder_final_improved_and_ssim.ipynb, which contains corresponding exploratory iterations and Exploratory Data Analysis (EDA) implementations evaluating these improved metrics and architectures.