Neural Image Compression and Explanation: Submitted By: Sampad Mohanty 2002070059

Neural Image
Compression and
Explanation
Submitted by:
Sampad Mohanty
2002070059
01 OVERVIEW
TABLE OF CONTENTS
02 IMPORTANCE OF IMAGE COMPRESSION
03 NICE FRAMEWORK
04 RESOURCES
05 ROADMAP
06 EXPECTED RESULTS AND FUTURE WORK PLAN

OVERVIEW
▪ We make a system that explains and
compresses images
▪ The system makes a mask that shows

important and unimportant parts of the image
▪ It makes the image size smaller by blurring

the unimportant parts
▪ This keeps the image meaning and quality
▪ The system shall work well on different kinds

of images and tasks
IMPORTANCE OF IMAGE
COMPRESSION
▪ Reduce the computational cost and memory requirement of training and

inference on large and high-resolution images.
▪ Enhance the performance and robustness of image recognition and

analysis of models by removing noise and redundancy.
▪ Real-time video streaming, augmented reality, edge computing.
▪ Leverage the power and potential of DNN to learn from data how to
compress images in an optimal and adaptive way.
NICE
FRAMEWORK
THE HARD
CONCRETE SMOOTHNESS
GRADIENT REGULARIZATION
ESTIMATOR
SEMANTIC
IMAGE SPARSE MASK
COMPRESSION GENERATOR
RESOURCES
MNIST CIFAR10 Caltech DATASETS

Modified National Institute Canadian Institute For 101 categories of images
of Standards and Advanced Research. with 300 x 200 pixels.
Technology 32 x 32 pixels images. Approximately around
28 x 28 pixels 70000 50000 Training images and 800 images per category.
images 10000 test images.
Other
RTIS Saliency Map CAM METHODS
Real-Time Image It will be generated by Class Activation Mapping
Saliency gradient-based methods helps the idea of using the feature
measures the effect of us to understand what maps from the last
perturbing a training features a model uses for convolutional layer of a
point on the model’s prediction. CNN.
prediction
EXPECTED RESULTS
RTIS CAM Saliency Map NICE
MNIST -- -- -- Better
CIFAR10 -- -- -- Better
Caltech -- -- -- Better
Internship on
Hardware for
Machine Learning
Guided by:
Professor Bibhudatta Sahoo
Department of Electronics and Electrical Communication
Indian Institute of Technology, Kharagpur
01 ROADMAP
TABLE OF CONTENTS
02 MACHINE LEARNING and DEEP LEARNING
03 CONVOLUTIONAL NEURAL NETWORKS
04 YOLO and FaceNet
05 IMPLEMENTATION of CNN USING VERILOG
06 FUTURE WORK PLAN

ROADMAP
Deep CNN
Learning Algorithms
Machine Single Layer

CNN
Learning CNN in Verilog
Machine Learning
REGRESSION
It predicts the continuous output
variables based on the independent
input variable.
CLASSIFICATION
It predicts the class of the dataset
based on the independent input
variable. Class is the categorical or
discrete values.
Deep Learning
Optimization
Gradient Checking Methods
Regularization Verifies the Help avoid
correctness of the getting stuck in
Initialization Reduce the
gradients computed local minima or
Setting the initial complexity and saddle points.
by backpropagation.
values of the overfitting of a
parameters of a neural network.
neural network.
Convolutional Convolutional
Layer
Neural
Applies a set of filters to the
input data, such as an image,
and produces a set of feature
Networks
maps. (CNN)
Pooling Layer
Reduces the size and
complexity of the feature
maps produced by the
convolutional layer
Fully Connected Layer

Connects all the neurons in the
previous layer to all the neurons in the
current layer.
CNN Algorithms
ResNet YOLO
Uses residual blocks, which are 01 You Only Look Once
layers that add the input to the Object Detection and Localization
output of a subnetwork, to avoid the Divides the input image into a grid of
02
problem of vanishing or exploding cells and predicts the bounding
gradients in deep networks. boxes and class probabilities for
each cell.
03
U-Net FaceNet
Perform semantic segmentation and Mapping from face images to a
image restoration. An encoder that
downsamples the input image and
04 compact Euclidean space where
distances directly correspond to a
extracts features, and a decoder that measure of face similarity.
upsamples the features and
produces the output.
YOLO FaceNet
Car Detection for Face Recognition

Autonomous Driving for Surveillance
Implementation
in Verilog
Perform a 3x3 convolution dot product on a 7x7 image in both software

(Python) and hardware (Verilog) domain.
(Implementation of a Single Layer CNN in Verilog)

Future Work Plans
Ongoing Projects Minor Project Future Projects

User Authorization System through Neural Image Compression and Will try to extend it for further work
Face Recognition Explanation towards more accuracy and precision.
Internship
Certificate
THANK YOU

Neural Image Compression and Explanation: Submitted By: Sampad Mohanty 2002070059

Uploaded by

Copyright:

Available Formats

Neural Image Compression and Explanation: Submitted By: Sampad Mohanty 2002070059

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Neural Image Compression and Explanation: Submitted By: Sampad Mohanty 2002070059

Uploaded by

Copyright:

Available Formats

Neural Image

06 EXPECTED RESULTS AND FUTURE WORK PLAN

▪ The system makes a mask that shows

▪ It makes the image size smaller by blurring

▪ This keeps the image meaning and quality

▪ The system shall work well on different kinds

▪ Reduce the computational cost and memory requirement of training and

▪ Enhance the performance and robustness of image recognition and

▪ Real-time video streaming, augmented reality, edge computing.

MNIST CIFAR10 Caltech DATASETS

RTIS CAM Saliency Map NICE

03 CONVOLUTIONAL NEURAL NETWORKS

04 YOLO and FaceNet

05 IMPLEMENTATION of CNN USING VERILOG

06 FUTURE WORK PLAN

Machine Single Layer

Fully Connected Layer

Car Detection for Face Recognition

Perform a 3x3 convolution dot product on a 7x7 image in both software

(Implementation of a Single Layer CNN in Verilog)

Ongoing Projects Minor Project Future Projects

You might also like