Image Processing Advanced
Deep Learning and Applications
About This Level
The advanced level covers modern image processing centered on deep learning. Starting from image classification with CNNs, it moves through object detection, semantic segmentation, and image generation with GANs and diffusion models, covering many of the core techniques in modern image processing.
Learning Objectives
- Understand CNN architecture and its application to image classification
- Master object detection methods such as YOLO, SSD, and Faster R-CNN
- Learn segmentation methods such as U-Net and DeepLab
- Understand image generation with GANs, VAEs, and diffusion models
- Learn specialized applications such as medical image processing
Diagram: CNN Architecture
Table of Contents
1. CNN and Image Classification
Foundations of deep learning.
- Convolution and pooling layers
- LeNet, AlexNet, VGG
- ResNet, DenseNet
- Transfer learning and fine-tuning
2. Object Detection
Joint estimation of location and class.
- The R-CNN family
- The YOLO series (v3, v5, v8, v11)
- SSD, RetinaNet
- Anchor-free detectors
3. Semantic Segmentation
Pixel-level classification.
- FCN (Fully Convolutional Network)
- U-Net and encoder-decoder
- The DeepLab series
- Instance segmentation
4. Image Generation Models
Generating new images.
- VAE (variational autoencoder)
- GAN and DCGAN
- StyleGAN, BigGAN
- Diffusion models (DDPM, Stable Diffusion)
5. Vision Transformer
Applying the attention mechanism.
- ViT architecture
- Patch embedding
- CLIP, DINO
- Swin Transformer
6. Medical Image Processing
Applications to a specialized field.
- CT/MRI image preprocessing
- Organ segmentation
- Lesion detection
- 3D image processing
Diagram: Object Detection
Diagram: Image Generation (GAN)
Key Models and Loss Functions
Cross-Entropy Loss (Classification)
$\mathcal{L} = -\displaystyle\sum_{c=1}^{C} y_c \log(\hat{y}_c)$ ($C$: number of classes, $y$: ground truth, $\hat{y}$: prediction)
IoU (Intersection over Union)
An evaluation metric for detection and segmentation: $\text{IoU} = \dfrac{|A \cap B|}{|A \cup B|}$
The GAN Objective
$\min_G \max_D \mathbb{E}_{x}[\log D(x)] + \mathbb{E}_{z}[\log(1 - D(G(z)))]$
Applications You Can Understand at This Level
Autonomous Driving
Environment perception that integrates object detection, segmentation, and depth estimation.
Medical Diagnosis Support
Lesion detection from CT/MRI, organ segmentation, and pathology image analysis.
Image Generation and Editing
Text-to-image, inpainting, style transfer, and super-resolution.
Manufacturing Quality Inspection
Defect detection and automated visual inspection. Reject decisions via anomaly detection.
Supplementary Reading
Readings that build intuition for the essentials of modern image processing and deep learning, using diagrams (including 3D).
Total Variation (TV) Regularization and Denoising
Understand how TV regularization ("data fidelity + an L1 penalty on the gradient") preserves edges.
Compressed Sensing and Sparse Recovery
Understand the principle by which sparsity and L1 minimization recover a signal from few measurements.
Optimal Transport and Images
Understand optimal transport as minimizing the "cost of moving mass" between distributions, giving the Wasserstein distance.
Wavelet Transform and Multiresolution Analysis
Understand wavelets as a multiresolution decomposition local in both position and scale.
Residual Connections and the Loss Landscape
Understand how residual connections ease gradient flow and smooth the loss landscape to aid training.
Transposed Convolution and Upsampling
Understand transposed convolution as "stamping the kernel onto the input" to upsample.
Diffusion Model Noise Schedule
Understand a diffusion model as a "forward process that adds noise" and a "reverse process that predicts and removes it."
Self-Attention, Visualized
Understand self-attention as each token attending to all tokens with learned weights.
Depth Maps and 3D Point Clouds
Understand a depth map as per-pixel distance, convertible into a 3D point cloud by back-projection.
Object Detection Metrics: IoU, NMS, mAP
Understand how IoU, NMS, and mAP handle "match decision," "duplicate removal," and "overall evaluation."
Prerequisites
- Intermediate-level content (morphology, feature extraction, classical object detection)
- Foundations of machine learning (loss functions, optimization, overfitting)
- Basics of Python and PyTorch / TensorFlow
- Linear algebra, calculus, probability and statistics