Computer Vision : Latest Usecases and Innovations (as of Year 2023)

Some of the most important usecases of Computer Vision and Image Analytics are:

1. Image Classification
   • Use CNN model (Convolutional Neural Network) refer MNIST dataset here – http://yann.lecun.com/exdb/mnist/
   • Around 70,000 images (60k training and 10k test images) of handwritten Numbers (0 to 9)
   • Refer Explanation Here – https://www.tensorflow.org/datasets/catalog/mnist
2. Object Detection and Localization
   • We have multiple pre-trained models of the same, e.g. R-CNN, Fast R-CNN and Faster R-CNN
   • Also YOLO (You Only Look Once) model is quite famous for low resolution images
     • SSD (single-shot multibox detector) Model is also quite relevant for low resolution images, SSD300/500 Models, and MobileNet V2.
     • We can directly use these models for prediction by leveraging openCV library’s dnn (deep neural network) module
     • We can also use Transfer Learning to leverage these pre trained model’s Architecture and Weights on our custom images.
   • ImageNet – Get Real World Images from here – https://www.image-net.org/download.php
   • This is similar to WordNet (used in NLP)- https://wordnet.princeton.edu/
   • Kaggle Competition for the same – https://www.kaggle.com/c/imagenet-object-localization-challenge/overview/description
3. Transfer Learning and Fine Tuning
4. Auto Encoders
   • can be used for noise reduction in images (denoising)
5. Siamese Networks
   • Contains 2 similar networks, which tries to generate features and compute loss function around similarity check
   • To be used when we have less image data to train the system
6. GANs (Generative Adversarial Networks)
   • Generative AI for images
   • Contains 2 models – (Generator and Discriminator)
   • Some of the Usecases/Applications
     • DeepFake – https://deepfakesweb.com/
7. Advance Generative AI (Text To Image)
   • Google Imagen Model
     • https://imagen.research.google/
   • OpenAI’s ImageGPT (DALL-E 2) – https://labs.openai.com/
     • The name DALL-E is a blend of Salvador Dali, the famous artist, and Pixar’s WALL-E movie
   • Stability AI’s Diffusion Model (Stable Diffusion)
     • https://stability.ai/stablediffusion
   • Google Parti Model (Pathways Autoregressive Text-to-Image Model)
     • https://sites.research.google/parti/

Image Processing and Prediction Models/algorithms:

1. OpenCV Package for Pre-Processing
2. LeCNN and then CNN (by Yann LeCunn) for image classification and object detection
3. AlexNet
4. Inception and Xception Models
5. DCNN (Deep CNN)
6. GANs
7. ESRGANs
8. Image-GPTs (GPT 3+)
9. Diffusion Models