Anda di halaman 1dari 20

IMAGE RECOGNITION AND

CLASSIFICATION USING ML
KSHITIJ SHAH
CHETANA LOKHANDE
AGENDA
• Introduction
• Project Description
• System Design Methodology
• System Architecture
• Technologies Used
• Model Development
• CNN, Modified CNN, LeNet & ZFNet
• Model Evaluation and results
• Demo
• Conclusion, Implementations, Recommendation
PROJECT DESCRIPTION
• Problem Statement
• When a user sees an object he/she likes but don’t know what keyword to use when
searching.
• User Experience is a major part of any system.
• All big companies invest a lot of time in making the user experience better.
• Machine Learning is used for recommendations by companies like Amazon.
• Machine learning already used in the field of defense and healthcare to perform image
recognition and classification tasks reducing human error.
• However, it can also be used for enhancing the User Experience.
INTRODUCTION

• Aims at improving the user experience using Neural network concepts.


• The system is capable of object recognition and classification.
• It will return the corresponding keyword based on the image input to the system.
• It can be used most appropriately with e-commerce application.
• Five targeted categories: Keyboard, Mouse, Watch, Laptop & Mobile Phones
SYSTEM DESIGN
METHODOLOGY
SYSTEM ARCHITECTURE
Technologies Used:

PYTHON SPYDER TENSORFLOW KERAS


SYSTEM
ARCHITECTURE
MODEL DEVELOPMENT

• CNN
• Modified – CNN
• LeNet
• ZFNet
CONVOLUTION NEURAL NETWORK (CNN)
• Neural network tries replicating the behavior of the brain with each neuron representing a cell in the
brain.
• Convolution Neural Network is a representation of the visual cortex of the brain.
• It is more efficient than a Fully connected Network.
• Has four main types of layers each having their own purpose:
• Convolution Layer
• ReLU (Rectified Linear Unit)
• Pooling Layer
• Fully Connected Layer
Convolution Layer: Detects various features and thus
applies filters on the Input matrix.
• Steps in Convolution Layer:
• Line up the feature and the image patch.
• Multiply each image pixel by the
corresponding feature pixel.
• Add them up
• Divide by total number of pixels in feature.

ReLU(Rectified Linear Unit) Layer: Consists of a


transform function which only activates if the input is
above a certain quantity
Flatten Layer: Converts the matrix to a 2D array
Dense Layer: Gives us the probability of the object to
belong to a particular class.
Pooling Layer: It is mainly used to shrink the image stack
to a smaller size.
• Steps for pooling:
• Pick a Window size
• Pick a stride
• Walk your window across your filtered images.
• From each window, take the maximum value.

Fully Connected Layer:


• Steps of Fully Connected Layer:
• A single list is created of the filtered and shrunk
image matrix.
• Compare the input vector with the derived lists.
• Classify using the difference/matching value.
MODIFIED - CNN
LENET

• LeNet is an architecture variation of CNN


• Major components of LeNet include sparse convolution layers and
max-pooling
• The lower layers consist of alternative convolution and max-
pooling layers. The upper layers are fully connected and
correspond to a traditional hidden layer + logistic regression
• Trains on 60000 parameters and hence has a higher accuracy than
that of a simple CNN model
ZFNET
• Introduces a visualization technique
• Allows to observe the evolution of features during training and to
diagnose potential problems with the model
• Feature maps are passed through a linear function, pooled in 3x3
regions and normalized across feature maps to give 55 by 55
element feature maps
• Similar thing is repeated in layers 2,3,4,5
MODEL EVALUATION AND RESULTS
CNN LENET ZFNET

No. of layers 15 12 14

Input Dimension 128 28 150


(Pixels)

Accuracy(Percentage) 91 89 34

Validation 83 75 28
Accuracy(Percentage)
DATA

• Training Data
• Images = 2000 (500/category)
• Test Data
• Images = 500 (100/category)
DEMO
• Implemented three algorithms namely CNN, LeNET and ZFNet.
• With the existing dataset and output class requirement, CNN has
proven to be the best Algorithm.

CONCLUSION,
• Neural networks can be used in order to deal visual data in the
IMPLICATION & field of defense and healthcare to reduce human error.
RECOMMENDATION
• An API can be built in order to facilitate integration with existing
systems.
WHAT WE LEARNT FROM THIS PROJECT?

Dealing with problems


Neural Network
of overfitting and
Concepts
underfitting

How each parameter Python, Tensorflow and


affects the final output Kerass
THANK YOU

Anda mungkin juga menyukai