Day # 146 - Part I - Deep Learning techniques for Computer Vision applied to embedded systems

A very interesting Final Year Paper - Deep Learning techniques for Computer Vision applied to embedded systems

Key points I loved in this paper, Re-posted from the paper. Very Good ML training and learning paper. Excellent Work.

Machine Learning Problems

• Classification - Train from a labelled dataset, Classify new incoming data to the class it belongs to. SVM, Decision Trees, Neural Networks, K Nearest Neighbors. Works on Discrete values
• Clustering - Grouping data that share similar characteristics. Data not labelled. Maximum Distance between clusters, Minimize distance between points in identified cluster. K-Means, Hierarchical Clustering, DBScan
• Regression - Considers continuous variable as output. Map input function to continuous output variable
• PCA - Principal Component Analysis. Exploit Matrix Decomposition, Eigen Values to retain principal Eigen Vectors, Reduce dimension retaining critical components
• Artificial Neural Network - Feedforward because output goes to next layer. Fully Connected - Each neuron propagates the result of computation to next neuron in following layer. Feed Forward + Fully Connected = Multi Layer Perceptron

Key Layers of Neural Network Design

• Activation Functions to use in Each Layer
• Loss Function to minimise Overfitting
• Backpropagation Algorithm to find right weights (CNN)
• Backpropagation Algorithm uses Stochastic Gradient Descent to compute Learning Rate

Computer Vision Applications

• Image Classification - Assigning class / label based on pretrained classes. 
• Image Classification and Localisation - Finding most relevant object in given image and bounding box of the relevant object in given image
• Object Detection - Extract Relevant Object and their location
• Instance Segmentation - Creates Overlap of detected objects/contours from extracted image. 

Deep Architecture

R-CNN - Region CNN

• First step is identify regions
• Second Step use CNN for identification
• Not suitable for real time applications
• Fast R-CNN, Improvement of R-CNN

Yolo

• Single Neural Network Applied to entire image
• You Only Look Once
• Bounding box created with probabilities containing the object
• Uses Predefined Grid Cells

SSD

• Single Shot Multi Box Detector
• Speeds up processing by Eliminating RPN
• Feature maps extracted and Convolution filter is applied

For Real-time processing Yolo - 45 FPS, SSD 59 FPS. 

Happy Learning!!!