Module 03: Computer Vision

Module 03: Computer Vision

计算机视觉 - 系统学习 CNN 架构演进与视觉任务。

---

Module Overview

Attribute	Value
Prerequisites	Module 02
Duration	3-4 weeks
Notebooks	18+
Difficulty	⭐⭐⭐ Intermediate

---

Learning Objectives

• ✅ 理解卷积神经网络的原理
• ✅ 掌握经典 CNN 架构 (LeNet → ViT)
• ✅ 应用迁移学习和微调技术
• ✅ 使用 Grad-CAM 等可视化技术解释模型

---

Architecture Evolution

LeNet (1998)     →    AlexNet (2012)    →    VGG (2014)
    5 layers            8 layers            16-19 layers
    60K params          60M params          138M params

        ↓                   ↓                   ↓

GoogLeNet (2014) →    ResNet (2015)     →    DenseNet (2016)
   Inception              Skip connections      Dense connections
   22 layers              50-152 layers         121-264 layers

        ↓                   ↓                   ↓

EfficientNet (2019) → Vision Transformer (2020) → Swin (2021)
   Compound scaling       Self-attention           Shifted windows
   5.3M-66M params        86M-632M params          88M-197M params

---

Submodules

01. CNN Basics

Topic	Key Concepts	Notebooks
Convolution	Filters, Stride, Padding	01_convolution.ipynb
Pooling	Max, Average, Global	02_pooling.ipynb
Architecture Design	Feature Extraction + Classifier	03_architecture.ipynb

Key Equations:

Convolution:
(I * K)ᵢⱼ = Σₘ Σₙ I(i+m, j+n) · K(m, n)

Output Size:
O = (I - K + 2P) / S + 1
where I=input, K=kernel, P=padding, S=stride

02. Classic Architectures

Architecture	Key Innovation	Notebooks
LeNet-5	First successful CNN	01_lenet.ipynb
AlexNet	ReLU, Dropout, GPU	02_alexnet.ipynb
VGG	Small filters (3×3)	03_vgg.ipynb
GoogLeNet	Inception modules	04_googlenet.ipynb
ResNet	Residual connections	05_resnet.ipynb
DenseNet	Dense connections	06_densenet.ipynb
EfficientNet	Compound scaling	07_efficientnet.ipynb

03. Transfer Learning

Technique	Description	Notebooks
Feature Extraction	Freeze backbone	01_feature_extraction.ipynb
Fine-tuning	Unfreeze layers	02_fine_tuning.ipynb
Progressive Resizing	Curriculum learning	03_progressive.ipynb

Example:

base_model = keras.applications.ResNet50(
    weights='imagenet',
    include_top=False,
    input_shape=(224, 224, 3)
)
base_model.trainable = False

model = keras.Sequential([
    base_model,
    keras.layers.GlobalAveragePooling2D(),
    keras.layers.Dense(256, activation='relu'),
    keras.layers.Dropout(0.5),
    keras.layers.Dense(num_classes, activation='softmax')
])

04. Model Visualization

Technique	Purpose	Notebooks
Grad-CAM	Class activation maps	01_gradcam.ipynb
Occlusion Sensitivity	Region importance	02_occlusion.ipynb
Filter Visualization	Feature detectors	03_filters.ipynb

---

Benchmark Results

Model	ImageNet Top-1	Params	FLOPs
ResNet-50	76.1%	25.6M	4.1G
ResNet-152	78.3%	60.2M	11.6G
EfficientNet-B0	77.1%	5.3M	0.4G
EfficientNet-B7	84.3%	66M	37G
ViT-B/16	77.9%	86M	17.6G
ViT-L/16	79.7%	304M	61.6G

---

← Module 02 | Module 04 →