Skip to content

CNNs and Computer Vision

Convolutional Neural Networks exploit spatial structure in images through local pattern detection and translation invariance. From classification to generation, CNNs revolutionized visual understanding.

Image as Data

Image = 3D tensor (height x width x channels). Grayscale: 1 channel. RGB: 3 channels. Each pixel = integer [0, 255].

For simple models: flatten to 1D (28x28 -> 784 features). Problem: loses spatial structure. CNNs solve this.

Convolution Layer

Slide small filter (kernel) across input. Each filter detects one pattern (edge, texture, shape).

  • Kernel size: 3x3, 5x5 typical. Small kernels stacked = large receptive field
  • Stride: step size. Stride 2 halves spatial dimensions
  • Padding: "same" preserves size, "valid" reduces
  • Channels: input channels matched by filter depth; output channels = number of filters
  • 1x1 convolution: linear combination of channels (bottleneck, dimensionality reduction)

Pooling

Reduce spatial dimensions. Max pooling (take max in window) most common. Typical: 2x2, stride 2.

CNN Pattern

Input -> [Conv -> BN -> ReLU -> (Pool)] x N -> Flatten -> Dense -> Output

Architecture Evolution

Architecture Year Key Innovation
AlexNet 2012 First deep CNN to win ImageNet. ReLU, dropout, data augmentation
VGG 2014 Only 3x3 convs stacked deeply. Simple, uniform
GoogLeNet/Inception 2014 Parallel convs of different sizes, concatenated. 1x1 bottlenecks
ResNet 2015 Skip connections: output = F(x) + x. 50-152 layers
DenseNet 2017 Each layer connected to all previous layers
MobileNet 2017 Depthwise separable convolutions for mobile/edge
EfficientNet 2019 Compound scaling (width, depth, resolution)
ViT 2020 Vision Transformer - apply transformer to image patches

ResNet Skip Connections

Solve vanishing gradient for deep networks: 

input -> Conv -> BN -> ReLU -> Conv -> BN -> (+input) -> ReLU
The identity shortcut lets gradients flow directly through the network.

Transfer Learning

Use pre-trained model (ImageNet: 1.2M images, 1000 classes) as starting point.

import torchvision.models as models
import torch.nn as nn

model = models.resnet50(pretrained=True)

# Option A: Feature extraction (freeze everything, replace head)
for param in model.parameters():
    param.requires_grad = False
model.fc = nn.Linear(2048, num_classes)

# Option B: Fine-tune last layers
for param in model.layer4.parameters():
    param.requires_grad = True
model.fc = nn.Linear(2048, num_classes)

Rule of thumb: smaller dataset = freeze more layers; larger dataset = fine-tune more.

Object Detection

Locate objects with bounding boxes + class labels.

Two-Stage Detectors

  • R-CNN: region proposals (selective search) -> CNN per region -> classify
  • Fast R-CNN: CNN on full image, extract features per region from feature map
  • Faster R-CNN: learned Region Proposal Network (RPN). Fully end-to-end

One-Stage Detectors

  • YOLO: divide image into SxS grid, each cell predicts boxes + classes. Single forward pass. Real-time
  • SSD: multi-scale feature maps for objects at different sizes

Detection Metrics

  • IoU: intersection / union of predicted and ground truth box. >= 0.5 = correct
  • mAP@0.5: mean Average Precision at IoU threshold 0.5
  • mAP@0.5:0.95: averaged over IoU thresholds 0.5 to 0.95 (step 0.05)
  • NMS (Non-Maximum Suppression): remove overlapping detections, keep highest confidence

Segmentation

Semantic Segmentation

Classify every pixel. No instance distinction. - FCN: replace FC layers with convolutions, upsample back - U-Net: encoder-decoder with skip connections. Excellent for medical imaging - DeepLab: atrous (dilated) convolutions for larger receptive field

Instance Segmentation

Detect objects AND segment pixel boundaries. - Mask R-CNN: Faster R-CNN + mask prediction branch per detected box

Segmentation Metrics

  • mIoU: mean IoU across classes (primary metric)
  • Dice coefficient: 2*|A intersect B| / (|A| + |B|). Equivalent to F1

Generative Models

  • GAN: Generator vs Discriminator adversarial training. Challenges: mode collapse, instability
  • VAE: encode to latent distribution, sample, decode. Smooth interpolation
  • Diffusion: iteratively denoise from pure noise. Current state-of-the-art for image quality
  • CycleGAN: unpaired image-to-image translation (style transfer, domain adaptation)

3D Vision

  • Depth estimation: predict distance per pixel from 2D image
  • Point cloud processing: PointNet for 3D point data
  • NeRF: learn 3D scene from 2D images, render novel views

Gotchas

  • Always normalize images (ImageNet stats: mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
  • Data augmentation is almost always beneficial for vision tasks
  • Larger input resolution = better accuracy but quadratic compute cost
  • Pre-trained models expect specific input sizes and normalization
  • YOLO versions vary widely - check which variant for fair comparison

See Also