🔍 Survey on Pretrained Model Performance for Seismic Image Denoising

This project conducts an in-depth study of how various pretrained image classification models perform when used as encoders for seismic image denoising tasks. Our architecture follows an encoder-decoder structure where:

• Encoder = pretrained model (e.g., EfficientNet, DenseNet, etc.)
• Decoder = UNet++, chosen for its dense skip connections and superior performance in preserving structural information during segmentation tasks.

---

🧠 Problem Statement

Seismic images often contain noise that can obscure important geological structures. Our objective is to:

Denoise seismic images using a hybrid deep learning pipeline that leverages powerful pretrained encoders with a UNet++ decoder.

---

🎯 Goals

• Evaluate the performance of state-of-the-art pretrained models as encoders on seismic image denoising.
• Analyze how different encoder backbones affect the structure preservation capability of UNet++.
• Compare performances to identify top-performing models.

---

📌 Methodology

1. Task: Image Denoising (Preserving key structures)
2. Model Architecture: Encoder–Decoder
   • Encoder: Pretrained classification model (from torchvision or timm)
   • Decoder: UNet++
3. Dataset: Seismic image dataset from Think-Towards Challenge
4. Loss Functions: Binary Cross Entropy + Dice Loss
5. Evaluation Metrics: PSNR, SSIM, IoU

---

🧱 Models Evaluated

Model	Year	Type
AlexNet	2012	CNN
VGG	2015	CNN
ResNet	2015	CNN
DenseNet	2016	CNN
ShuffleNet V2	2016	CNN
MobileNet V1/V2	2017	CNN
EfficientNet	2018	CNN
EfficientNet V2	2019	CNN
RegNet	2020	CNN
ConvNeXt	2021	CNN
RexNet	2020	CNN
MaxViT, Swin, ViT	2020–2022	❌ Not tested (resource constraints)

---

🏆 Key Findings

• ✅ Top Performers:

  • RexNet
  • DenseNet
  • EfficientNetV2

• ❌ Transformers (e.g., Swin, ViT, MaxViT) were excluded due to computational limitations.

---

💡 Why UNet++?

We chose UNet++ for its dense connections that enhance gradient flow and structural fidelity—ideal for segmentation-like tasks such as denoising, especially when retaining the underlying patterns in seismic data is crucial.

---

📊 Results & Visuals

Based on our experiments, the following encoder backbones performed best when paired with the UNet++ decoder for seismic image denoising:

• ✅ RexNet
• ✅ DenseNet
• ✅ EfficientNetV2

📸 Visual Results of Top Performers

Model	Sample Output
RexNet	(
)
DenseNet	(
)
EfficientNetV2	(
)

📑 View Full Slide Deck with All Experiments (PDF)