Skip to content

DNCNN_VERSIONS_COMPARISON_REPORT.md

Latest commit

 

History

History
57 lines (41 loc) · 4.38 KB

DNCNN_VERSIONS_COMPARISON_REPORT.md

File metadata and controls

57 lines (41 loc) · 4.38 KB

DnCNN-ECA Version Comparison Report

This report summarizes the architectural evolution of the public DnCNN-Lite ECA variants, from the earliest CPU-safe versions to the later V4 and V5 scripts, and links that evolution to both denoising metrics and downstream classification behavior.

1. Architecture Evolution and Features

File Version Attention Type Pooling Spatial Attention Config Style Main Characteristics
FIXED.py (V1) Simple ECA (fixed kernel) Global Average Pooling No Hardcoded Minimal CPU-safe ECA integration after convolution blocks.
PATCHED.py / v2.py Simple ECA (fixed kernel) Global Average Pooling Yes (SpatialLiteAttention) Hardcoded Added lightweight spatial attention to model not only which channels matter, but also where to focus spatially.
..._v3.py (fixed-patched branch) Squeeze-and-Excitation (SE) Global Average Pooling Yes Hardcoded Experimental stage where ECA was temporarily replaced by an SE block with linear layers.
FULL_PATCH_v3.py Advanced ECA Global Average Pooling Yes Config class Major refactor introducing DnCNNLiteECAConfig, temperature/gain controls, and mixed-precision support.
espi_dncnn_lite_eca_FULL_PATCH_v4.py Advanced ECA Global Average Pooling Yes Config class Stable thesis-era version with cleaner I/O, fair ECA vs no-ECA controls, and better experiment reliability.
espi_dncnn_lite_eca_FULL_PATCH_v5.py Extended ECA Dual Pooling (avg + max) Yes Config class Research-oriented extension with dual pooling, optional learnable temp/gain, multi-scale kernels, and placement presets.

2. Denoising Metrics

The values below summarize comparative evaluation on both synthetic validation and real ESPI pairs / averages.

Model Training Regime Val PSNR (Synthetic) Val SSIM Val EdgeF1 Real PSNR
V4 Baseline (NoECA) Pseudo-noisy synthetic supervision 27.24 dB 0.7846 0.7686 34.58 dB
V4 ECA Pseudo-noisy synthetic supervision 27.47 dB 0.7972 0.7787 34.50 dB
V5 Baseline (NoECA) Pseudo-noisy synthetic supervision 27.24 dB 0.7846 0.7686 34.58 dB
V5 ECA (Advanced) Pseudo-noisy synthetic supervision 27.24 dB 0.7852 0.7712 34.22 dB
V4R Baseline (NoECA) Real pairs (23,891 images) N/A N/A N/A 23.76 dB (real validation)
V4R ECA Real pairs (23,891 images) N/A N/A N/A 23.85 dB (real validation)

Key denoising observations

  • V4 ECA is clearly strongest on the reported synthetic denoising metrics, with gains of +0.23 dB PSNR and +0.0126 SSIM over the V4 baseline.
  • V5 ECA, despite being architecturally more ambitious, does not consistently outperform the simpler V4 ECA configuration in the reported experiments.
  • On real-aligned training pairs, adding ECA in the V4R regime provides a small but stable denoising gain over the no-ECA counterpart.

3. Downstream Classification Task

The table below summarizes how each denoising regime affected the downstream 5-class ResNet-18 classifier.

Pre-processing Pipeline Denoiser Training Data ECA Enabled Classification Accuracy (%) Classification Macro-F1 (%) dAcc vs Raw
No denoising (Raw) None No 97.70 93.99 0.00
V4 denoised Pseudo-noisy synthetic supervision (243 images) No 96.39 89.06 -1.31
V4 denoised Pseudo-noisy synthetic supervision (243 images) Yes 94.77 84.21 -2.93
V4R denoised (real-trained) Real pairs (23,891 images) No 98.76 96.07 +1.06
V4R denoised (real-trained) Real pairs (23,891 images) Yes (V4 ECA) 98.87 96.64 +1.17

Key downstream observations

  1. Training regime dominates architecture alone. The denoisers trained on pseudo-noisy synthetic data degraded downstream classification, even when image-quality metrics looked competitive.
  2. Real-aligned training is the decisive factor for downstream benefit. The real-trained V4R models improved the classification pipeline over the raw baseline.
  3. V4R ECA is the best overall system-level result. It achieved the strongest reported downstream Accuracy and Macro-F1 in the final thesis package.