Table of Contents
Preface
Acknowledgements
Author biography
1 Mathematical tools for computer vision
1.1 Probability, entropy and Kullback–Leibler divergence
1.1.1 Probability and Shannon entropy
1.1.2 Kullback–Leibler divergence and cross entropy
1.1.3 Conditional probability and joint entropies
1.1.4 Jensen’s inequality
1.1.5 Maximum likelihood estimation and over fitting
1.1.6 Application of expectation-maximization algorithm to find a PDF
1.2 Using a gradient descent algorithm for linear regression
1.3 Automatic gradient calculations and learning rate schedulers
1.4 Dataset, dataloader, GPU and models saving
1.5 Activation functions for nonlinear regressions
References
2 Image classifications by convolutional neural networks
2.1 Classification of hand written digits in the MNIST database
2.2 Mathematical operations of a convolution
2.3 Using ResNet9 for CIFAR-10 classification
2.4 Transfer learning with ResNet for STL-10 dataset
References
3 Image generation by GANs
3.1 The GAN theory
3.1.1 Implement a GAN for quadratic curve generation
3.1.2 Using a GAN with two fully connected layers to generate MINST Images
3.2 Applications of deep convolutional GANs
3.2.1 Mathematical operations of ConvTranspose2D
3.2.2 Applications of a DCGAN for MNIST and fashion MNIST
3.2.3 Using a DCGAN to generate fake anime-faces and fake CelebA images
3.3 Conditional deep convolutional GANs
3.3.1 Applications of a cDCGAN to MNIST and fashion MNIST datasets
3.3.2 Applications of a cDCGAN to generate fake Rock Paper Scissors images
References
4 Image generation by WGANs with gradient penalty
4.1 Using a WGAN or a WGAN-GP for generation of fake quadratic curves
4.2 Using a WGAN-GP for Fashion MNIST
4.3 WGAN-GP for CelebA dataset and Anime Face dataset
4.4 Implement of a cWGAN-GP for Rock Paper Scissors dataset
References
5 Image generation by VAEs
5.1 VAE and beta-VAE
5.2 Application of beta-VAE for fake quadratic curves
5.3 Application of beta-VAE for the MNIST dataset
5.4 Using VAE-GAN for fake images of MNIST and Fashion MNIST
References
6 Image generation by infoGANs
6.1 Using infoGAN to generate quadratic curves
6.2 Implementation of infoGAN for the MNIST dataset
6.3 infoGAN for fake Anime-face dataset images
6.4 Implementation of infoGAN to the rock paper scissors dataset
Reference
7 Object detection by YOLOv1/YOLOv3 models
7.1 Bounding boxes of Pascal VOC database for YOLOv1
7.2 Encode VOC images with bounding boxes for YOLOv1
7.2.1 VOC image augmentations with bounding boxes
7.2.2 Encoding bounding boxes to grid cells for YOLOv1 model training
7.2.3 Chess pieces dataset from Roboflow
7.3 ResNet18 model, IOU and a loss function
7.3.1 Using ResNet18 to replace YOLOv1 model
7.3.2 Intersection over union (IOU)
7.3.3 Loss function
7.4 Utility functions for model training
7.5 Applications of YOLOv3 for real-time object detection
References
8 YOLOv7 and YOLOv8 models
8.1 YOLOv7 for object detection for a custom dataset: MNIST4yolo
8.2 YOLOv7 for instance segmentation
8.3 Using YOLOv7 for human pose estimation (key point detection)
8.4 Applications of YOLOv8 models
8.4.1 Image object detection, segmentation, classification and pose estimation
8.4.2 Object counting on an image or a video frame
8.4.3 Car tracking and counting for a video file
8.4.4 Fine tuning YOLOv8 for objection detection and annotation of a custom dataset
8.5 Using YOLO-NAS models for object detection
References
9 U-Nets for image segmentation and diffusion models for image generation
9.1 Retinal vessel segmentation by a U-Net for DRIVE dataset
9.2 Using an attention U-Net diffusion model for quadratic curve generation
9.2.1 The forward process in a DDPM
9.2.2 The backward process in the DDPM
9.3 Using a pre-trained U-Net from Hugging Face to generate images
9.4 Generate photorealistic images from text prompts by stable diffusion
References
10 Applications of vision transformers
10.1 The architecture of a basic ViT model
10.2 Hugging Face ViT for CIFAR10 image classification
10.3 Zero shot image classification by OpenAI CLIP
10.4 Zero shot object detection by Hugging Face’s OWL-ViT
References
11 Knowledge distillation, DINO, SAM, MiDaS and NeRF
11.1 Knowledge distillation for neural network compression
11.2 DINO: emerging properties in self-supervised vision transformers
11.3 DINOv2 for image retrieval, classification and feature visualization
11.4 Segment anything model: SAM and FastSAM
11.5 MiDaS for image depth estimation
11.6 Neural radiance fields for synthesis of 3D scenes
11.6.1 Camera intrinsic and extrinsic matrices
11.6.2 Using MLP with Gaussian Fourier feature mapping to reconstruct images
11.6.3 The physics principle of render volume density in NeRF
References
Appendix
