Deep Learning for Image Analysis
- 1. Deep Learning for Image Analysis Levan Tsinadze PulsarAI Deep Learning Tbilisi
- 2. Artificial Neural Network Neurons Weights Graph
- 3. Neuron
- 4. Specifications Input Layer – Hidden Layers – Output Layer “Easy” Functions Parallelism (Matrices)
- 5. ANN / DNN
- 6. Problems in CV Object recognition Object detection Semantic segmentation
- 10. Hybrid Models – Detect - Recognize
- 11. Input Image RGB
- 13. Input Image as Tensor
- 14. Convolutional NN Image as 3 dimensional array – tensor of pixels Filters Small matrix of weights slide on image (tensor) Step by step extract features from image Locall connectivity and spatial invariance
- 16. CNN
- 17. Filter - Input - Output Input – tensor Output – matrix Depth is shrinked
- 18. Convolutional Block Convolutional fiters (often more than one) Activation Pooling
- 19. Features by Blocks - Hierarchy
- 20. Number of Filters Input channel – Image Depth (RGB) Output channel – Hyper Parameter number of convolutional filters (kernels) Feature Map – Generated Matrix for Each Patch Feature Maps on top of each other
- 22. Many to One
- 23. Weight Shearing Features map – small amount of weights 6x6 input 10 features matrices 3x3 each 90 parameters produces 4x4x10 tensor 224x224x3 to 2048 get 101054464 parameters Fully connected layer 6x6x4x4 = 576 32 Features map 256 or even 512 or 1024 features map Still better presision with less parameters = less calculation
- 24. FP – Sparce FFNN
- 25. Activations ConvLayer Activation (Elementwise) ConvLayer Etc Tensor → Tensor → Tensor → ...
- 26. ReLU x = x < 0 ? 0 : x or x = max(0, x) 9:15 – Graph Df/Dx = 0 when x = 0 Easy to Propagate (Forward and Back) Easy to See Results
- 27. ReLU - Watch
- 28. ReLU - Graph
- 29. Pooling
- 30. Pooling Types Max-Pooling Average – Pooling L2-Norm-Poolig
- 32. Pooling only on Matrices Only Matrices Pooling for each Channel Same Dimensions Height Changes Width Changes
- 33. Dropout
- 34. MNIST
- 35. LeNet Convolutional Layer Max pooling layer Convolutional layer Max pooling layer Dropout (if training) Fully connected layer Dropout (if training) Fully connected layer
- 36. Training Epochs Train Shuffle Train again Only on training set Avoid local minimum
- 37. PyTorch Dynamic execution Imperative Object-oriented http://pytorch.org/
- 39. Fine-Tuning / Transfer Learning Freeze lower layers Retrain higher layers Small dataset Fast training Less resources
- 40. ResNet-18 Residual connections Vanishing gradient Very deep convolutional neural network Easy architecture Eveilable pretrained (On ImageNet) weights
- 43. One / N – Shot Learning Retraining problems Labeling problems (Letters OK, Numbers OK Faces NOT OK) Features extractors Feature Search
- 45. Features Extractor Text Image Features - Vector
- 46. Faces