Deep learning with_computer_vision
0 likes143 views
This document outlines the course contents for a seminar on deep learning with computer vision. The 3-day course will cover topics such as convolutional neural networks, generative models, object detection, semantic segmentation, and applications of computer vision like image captioning, retrieval, and generation. Students will learn popular CNN architectures, train models using TensorFlow and Keras, and develop real-world computer vision applications. Hands-on exercises and case studies are included to provide practical experience with computer vision tasks.
Deep learning with_computer_vision
- 1. Deep Learning with Computer Vision SEMINAR mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com ☛ Description ☛ Key Skills ☛ Prerequisites ☛ Instructional Method ☛ course contents contents
- 2. Deep Learning with Computer Vision SEMINAR mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com ☛ Software Tools Day1 ☛ Convolutional Neural Networks Internals ☛ Deep Generative Models Day2 ☛ OpenCV ☛ Deep learning for computer vision Day3 course contents
- 3. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com Description: ■ You’ve probably heardthat DeepLearning is making news acrosstheworld as one of the most promisingtechniquesin machine learning,especially foranalyzing image data.With everyindustrydedicatingresourcesto unlockthe deep learning potential, to be competitive,youwillwant to use thesemodels in tasks such asimage tagging, object recognition,speechrecognition,andtext analysis.In this trainingsession you will build deep learningmodelsforComputerVision. One the detailed case study will be using attention basedmechanismto do image segmentationandobject recognition.Anotherdetailed case study would be image captioning and understandingusinga combinationattentionbasedCNNand sequence based model(LSTM). Key Skills: ■ face recognition ■ image generation ■ video classification ■ image captioning ■ medical image segmentation ■ productsearch ■ OpticalCharacter/Word/Sentence Recognition ■ persondetection Prerequisites: ■ This is an advancedlevelsession andit assumes that youhavegoodfamiliarity with Machine learning. ■ Working Knowledgeofpython ■ Machine LearningInternals ■ All sequence based models like RNN, LSTM, GRUs, Attention,Language Models must be known as in "Modern NaturalLanguage Processing(NLP)with Deep Learning session(Kindlyreferto the first itemin the briefTOC)". Instructional Method: ■ This is an instructorled course provideslecture topicsandthe practicalapplication of Deep Learning and theunderlying technologies.It pictorially presentsmost concepts and there is a detailed case studythat stringstogetherthe technologies,patternsand design.
- 4. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com Deep Learning with Computer Vision ■ Software Tools ■ Tensorflow • Installation • Sharing Variables • Creating Your First Graph and Running It in a Session • Managing Graphs • Visualizing the Graph and Training Curves Using TensorBoard • Implementing Gradient Descent • Lifecycle of a Node Value • Linear Regression with TensorFlow • Modularity • Saving and Restoring Models • Name Scopes • Feeding Data to the Training Algorithm ■ Keras ■ Convolutional Neural Networks Internals • Pooling layer • Image augmentation • Convolutional layer • History of CNNs • Convolutional layers in Keras • Code for visualizing an image • Input layer • How do computers interpret images? • Practical example image classification • Convolutional neural networks • Dropout ■ Attention Mechanism for CNN and Visual Models • Types of Attention • Glimpse Sensor in code
- 5. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • Attention mechanism for image captioning • Hard Attention • Applying the RAM on a noisy MNIST sample • Recurrent models of visual attention • Using attention to improve visual models • Reasons for sub-optimal performance of visual CNN models • Soft Attention ■ Build Your First CNN and Performance Optimization • Convolution and pooling operations in TensorFlow • Convolutional operations • Using tanh • Convolution operations in TensorFlow • Regularization • Fully connected layer • Weight and bias initialization • Pooling, stride, and padding operations • CNN architectures and drawbacks of DNNs • Applying pooling operations in TensorFlow • Using sigmoid • Training a CNN • Using ReLU • Activation functions ■ Building, training, and evaluating our first CNN • Creating a CNN model • Defining CNN hyperparameters • Model evaluation • Dataset description • Loading the required packages • Running the TensorFlow graph to train the CNN model • Preparingthe TensorFlow graph • Loading the training/test images to generate train/test set • Constructing the CNN layers
- 6. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com ■ Model performance optimization • Applying dropout operations with TensorFlow • Building the second CNN by putting everything together • Appropriate layer placement • Which optimizer to use? • Creating the CNN model • Dataset description and preprocessing • Number of neurons per hidden layer • Number of hidden layers • Batch normalization • Memory tuning • Training and evaluating the network • Advanced regularization and avoiding overfitting ■ Popular CNN Model Architectures • Architecture insights • ResNet architecture • AlexNet architecture • VGG image classification code example • Introduction to ImageNet • VGGNet architecture • GoogLeNet architecture • LeNet • Traffic sign classifiers using AlexNet • Inception module ■ Transfer Learning • Multi-task learning • Target dataset is small but different from the original training dataset • Autoencoders for CNN • Applications • Target dataset is large and similar to the original training dataset • Introducing to autoencoders
- 7. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • Convolutional autoencoder • Target dataset is large and different from the original training dataset • Transfer learning example • Feature extraction approach • Target dataset is small and is similar to the original training dataset • An example of compression ■ GAN: Generating New Images with CNN • Feature matching • GAN code example • Deep convolutional GAN • Adding the optimizer • Training a GAN model • Semi-supervised learning and GAN • Pixpix - Image-to-Image translation GAN • Calculating loss • Semi-supervised classification using a GAN example • CycleGAN • Batch normalization ■ Object Detection and Instance Segmentation with CNN • Creating the environment • Fast R-CNN (fast region-based CNN) • The differences between object detection and image classification • Mask R-CNN (Instance segmentation with CNN) • Cascading classifiers • Haar Features • Faster R-CNN (faster region proposal network-based CNN) • Traditional, nonCNN approaches to object detection • R-CNN (Regions with CNN features) • Running the pre-trained model on the COCO dataset • Why is object detection much more challenging than image
- 8. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com classification? • The Viola-Jones algorithm • Preparingthe COCO dataset folder structure • Downloading and installing the COCO API and detectron library (OS shell commands) • Instance segmentation in code • Haar features, cascading classifiers, and the Viola-Jones algorithm • Installing Python dependencies (Python environment) ■ Popular CNN Model Architectures • Introduction to ImageNet • VGG image classification code example • GoogLeNet architecture • Architecture insights • Inception module • AlexNet architecture • VGGNet architecture • LeNet • ResNet architecture • Traffic sign classifiers using AlexNet ■ Deep Generative Models • Deep Boltzmann Machines • Back-Propagation through Random Operations • Restricted Boltzmann Machines • Generative Stochastic Networks • Boltzmann Machines for Structured or Sequential Outputs • Boltzmann Machines • Other Boltzmann Machines • Other Generation Schemes • Directed Generative Nets • Boltzmann Machines for Real-Valued Data • Evaluating Generative Models • Drawing Samples from Autoencoders
- 9. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • Deep Belief Networks • Convolutional Boltzmann Machines ■ OpenCV • The Core Functionality (core module) • Introduction to OpenCV • Object Detection (objdetect module) • Image Processing(imgproc module) • Deep Neural Networks (dnn module) • GPU-Accelerated Computer Vision (cuda module) ■ Deep learning for computer vision ■ Similarity learning ■ Human face analysis ■ Face landmarks and attributes • Multi-Task Facial Landmark (MTFL) dataset • The Kaggle keypoint dataset • The Multi-Attribute Facial Landmark (MAFL) dataset • Learning the facial key points ■ Face recognition • Finding the optimum threshold • The YouTube faces dataset • The labeled faces in the wild (LFW) dataset • The CelebFaces Attributes dataset • CASIA web face database • The VGGFace2 dataset • Computing the similarity between faces ■ Face detection ■ Face clustering ■ Algorithms for similarity learning • Visual recommendation systems • DeepRank • FaceNet
- 10. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • The DeepNet model • Contrastive loss • Triplet loss • Siamese networks ■ Classification ■ Image Classification ■ The bigger deep learning models • The DenseNet model • The Google Inception-V3 model • The VGG-16 model • The SqueezeNet model • The AlexNet model • Spatial transformer networks • The Microsoft ResNet-50 model ■ Other popular image testing datasets • The Fashion-MNIST dataset • The CIFAR dataset • The ImageNet dataset and competition ■ Training the MNIST model in TensorFlow • The MNIST datasets • Building a multilayer convolutional network • Building a perceptron • Loading the MNIST data ■ Training a model for binary classification • Transfer learning or fine-tuning of a model • Preparingthe data • Augmenting the dataset • Benchmarkingwith simple CNN • Fine-tuning several layers in deep learning ■ Developing real-worldapplications • Brand safety
- 11. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • Tackling the underfitting and overfitting scenarios • Gender and age detection from face • Choosing the right model • Fine-tuning apparel models ■ Image Retrieval ■ Model inference • Serving the trained model • Exporting a model ■ Understanding visual features • Embedding visualization • The DeepDream • Visualizing activation of deep learning models • Adversarial examples • Guided backpropagation ■ Content-based image retrieval • Matching faster using approximate nearest neighbour • Extracting bottleneck features for an image • Computing similarity between query image and target database • Autoencoders of raw images • Building the retrieval pipeline • Efficient retrieval • Advantages of ANNOY • Denoising using autoencoders ■ Generative models ■ Generative Adversarial Networks • Drawbacks of GAN • Image translation • InfoGAN • Conditional GAN • Adversarial loss
- 12. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • Vanilla GAN ■ Applications of generative models • Inpainting • Super-resolution of images • Blending • 3D models from photos • Text to image generation • Transformingattributes • Creating training data • Image to image translation • Interactive image generation • Artistic style transfer • Creating new animation characters • Predicting the next frame in a video ■ Neural artistic style transfer • Style transfer • Style loss using the Gram matrix • Content loss ■ Visual dialogue model • Algorithm for VDM • Discriminator • Generator ■ Video analysis ■ Extending image-based approaches to videos • Captioning videos • Regressing the human pose • Generating videos • Tracking facial landmarks • Segmenting videos ■ Exploring video classification datasets • UCF101
- 13. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com • YouTube-8M • Other datasets ■ Understanding and classifying videos ■ Approaches for classifying videos • Using trajectory for classification • Multi-modal fusion • Using 3D convolution for temporallearning • Classifying videos over long periods • Fusing parallel CNN for video classification • Attending regions for classification • Streaming two CNN's for action recognition ■ Image captioning ■ Understanding natural language processing for image captioning • Expressing words in vector form • Training an embedding • Converting words to vectors ■ Implementing attention-based image captioning ■ Approaches for image captioning and related problems • Retrieving captions from images and images from captions • Creating captions using image ranking • Using attention network for captioning • Using multimodal metric space • Knowing when to look • Using a condition random field for linking image and text • Using RNN on CNN features to generate captions • Using RNN for captioning • Dense captioning ■ Understanding the problem and datasets ■ Detection or localization and segmentation ■ Object Detection
- 14. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com ■ Object detection API • Re-training object detection models • Data preparation for the Pet dataset • The YOLO object detection algorithm • Monitoring loss and accuracy using TensorBoard • Pre-trained models • Training the model • Object detection training pipeline • Training a pedestrian detection for a self-driving car ■ Detecting objects in an image ■ Localizing algorithms • Convolution implementation of sliding window • Combining regression with the sliding window • Thinking about localization as a regression problem • Applying regression to other problems • The scale-space concept • Localizing objects using sliding windows • Training a fully connected layer as a convolution layer ■ Detecting objects • Single shot multi-box detector • Regions of the convolutional neural network (R-CNN) • Fast R-CNN • Faster R-CNN ■ Exploring the datasets • Intersection over Union • ImageNet dataset • PASCAL VOC challenge • COCO object detection challenge • Evaluating datasets using metrics • The mean average precision ■ Semantic Segmentation
- 15. mobile:+91.7719882295 mailto:sales@anikatechnologies.com Website : www.anikatechnologies.com ■ Segmenting satellite images • Modeling FCN for segmentation ■ Datasets ■ Predicting pixels • Understandingthe earth from satellite imagery • Diagnosing medical images • Enabling robots to see ■ Algorithms for semantic segmentation • Large kernel matters • The Fully Convolutional Network • RefiNet • Upsampling the layers by pooling • The SegNet architecture • DeepLab • PSPnet • Skipping connections for better training • Sampling the layers by convolution • Dilated convolutions ■ Ultra-nerve segmentation ■ Segmenting instances