Transfer Learning for Image Classification

Transfer learning Remove

Predict

Train the model

Images / Datasets

Pretrained Mode

Alexnet

Googlenet

Resnet50

Etc

Softmax Layer

Classification layer

Fully Connected Layer

The Deep network (Feature Extraction)

Intial Learning rate

Mini batchsize

max Epochs

Hyperparameters

Categories

Score

Balance dataset

Partition

Datastore

Alot of dataset that

Different Input Size

Interpreting Network behavior

Activations

more information of interpretability

Feature Extraction

CN

Normal

FC

Pooling

Relu

ALEXNEt

First Layers are feature extraction
FC7 = 1x4096

ML

Tree

Logistics

SVM

Any other ML

Regression

VECTOR

LIME

visualize each activations

OcclusionSensitivity

Creating Networks

Land Cover Classification
4 Channel

B

G

NIR

R

28x28

Create/Training NN

Transfer Learning

Training From Scratch

Pretrained

Image for Inference

Labeled data (Image)

Pretrained Network

Replacing the last 3 layer

First layers -> weights / biases are
trained for Feature Extraction

Hyperparameters

Initial learning rate

Minibatch Size

Plot (training progress)

Epochs

Learning Dropout

Architecture

DAG

Series

Alexnet

resnet50

Trained with Image net

1000 categories

Solver

Training a Network

Plots / Loss Function

Hyper Parameters

Tuning XXX number parameters

Accuracy

Loss

CN

FC

Other

Initial learning rate

Dataset

Mini Batch Size

Scheduled Learning Rate

Fixed Learning Rate

Optimization

Training

Validation

Testing

Improve Accuracy

Architecture

Methods Hyperparamter
Tuning

Randomsearch

Bayesian Optimization

Gridsearch

Performing Image Regression

Classification

Regression

Replace

it is not ImageDatastore
but it is in a for of a Table

  1. Image Location
  2. Regression Values

Scoring

Color Correction

Presets for R

Presets for G

Presets for B

Image Stabilization

Correction Angle

Remove Classification Layer
Replace Regression Layer

Remove Fully Connect Layer
Replace new Fully Connected layer (OUTPUT is considered)

RSME

R2 / R2 adjusted

MAE

MAPE

Using Deep Learning for Computer Vision

Object Classification

Pixel Sementation

Object Localization

Image to Image Regression

Labels

Ground Truthlabler

ImageLabeler

Lidar Labeler

VideoLabeler

Yolov2 (YOLO)
You Only Look Once

Pretrained Network

Grid

Anchor

Thresholding

Look in Computer Vision
Toolbox Documentation

AUTOMATE
As much as Possible

Signal labeller

LSTM

Sequence -> Forecast (Future)

Sequence -> Sequence

Train from Scratch

Sequence Input Layer

LSTM Layer

State / Memory

Fully Connected Layer

Sequence -> Classification

click to edit

MATLAB -> EXPERIMENT MANAGER

Classification Layer

Audio Labeller