Semantic Image Analysis by Using Data Augmentation and Transfer Learning

Project Title

Semantic Image Analysis by Using Data Augmentation and Transfer Learning

Project Area of Specialization Artificial IntelligenceProject Summary

Executive Summary

This proposal aims to present a basic overview in the area of image classification based on Deep

Neural Networks (DNNs).

Introduction

1.1 Major Research Area

Computer Vision and Digital Image Processing

1.2 List of topics available for research in major area

 Deep learning

 Image Classification

 Image Retrieval

 Object Recognition

 Automatic Image Annotation

 Image Analysis

1.3 Selected topics for this research study

Image analysis based on deep learning

1.4 Introduction

This proposal aims to present a basic overview in the area of image classification based on Deep Neural Networks (DNNs). The automatic classification of digital images plays a vital role in different computer vision applications such as to recognize objects such as indoor, outdoor, streets, subways, street signs, traffic signals etc. Scene analysis, Computer-Aided Diagnosis (CAD), object recognition, image retrieval and image annotation are some of the applications that rely on the performance of a classification-based model. In last two decades, significant research is reported that is based on traditional visual features that can be either local or global (such as spatial layout, colour, texture, shape, interest point detectors or a combination of these). The traditional image classification and feature extraction techniques are usually domain specific and their performance degrades on large-scale dataset due to the use of handcrafted features and they lack generalization ability. Recent approaches for classification of images are based on the use of deep learning methods as they provide an effective and reliable framework that can compute the image labels. In this research proposal, we aim to propose a novel image representation that will be based on deep network features extracted through a pre- trained DNN on the basis of data-augmentation and transfer learning to sort out the best performance for the specific cases/standard datasets.

Project Objectives

In recent few years, due to the availability of large scale image datasets, the image classification approaches are shifted to the use of deep-learning models as they have significantly improved the classification accuracy for scene analysis as they are based on reliable layers for hidden network [3, 4]. AlexNet[5] , VGGNet[6] , and GoogLeNet[7] are the examples of most popular deep learning models that are considered as more discriminative when compared with handcrafted features (local and global) and unsupervised learning approaches[1-3]. Computations equipped with GPU units are considered as one of the basic requirements when dealing with deep learning-based approaches [8]. The main advantage of deep-learning approaches as compared to others is learning ability from the raw image through deep-network architecture [2]. The previous approaches require a lot of domain specific engineering while in deep learning approaches this is handled by the framework of complex network consisting of hidden layers. When compared with unsupervised learning approaches, deep learning methods are reported as more reliable as learning is based on multiple stacks of layers. The  training time/model learning and huge number of training samples are the two strong limitations for a deep-learning architecture [1-3]. The deep learning based framework has replaced the previous trends that were used for image classification and image analysis. The approach based on transfer learning is applied to a pre-trained deep neural network on a domain with different number of classes [1-3]. Lot of training samples is one of the basic requirements for a deep-learning architecture and if samples are not available, data augmentation is applied on images to enhance the learning process [1-3].

References

1. Alzu’bi A, Amira A, Ramzan N. Semantic content-based image retrieval: A

comprehensive study. Journal of Visual Communication and Image Representation.

2015;32:20–54.

2. Tousch AM, Herbin S, Audibert JY. Semantic hierarchies for image annotation: A

survey. Pattern Recognition. 2012;45(1):333–345.

3. Zhang D, Islam MM, Lu G. A review on automatic image annotation techniques. Pattern

Recognition. 2012;45(1):346–362.

Project Implementation Method

Proposed Implementation Scheme

The proposed study scheme is mentioned below with details of topics/sub-topic:

Image classification using machine learning approaches

Image classification using deep learning

Benefits of the Project

The main benefit and aim is to present a basic overview in the area of image classification based on Deep Neural Networks (DNNs). The automatic classification of digital images plays a vital role in different computer vision applications such as to recognize objects such as indoor, outdoor, streets, subways, street signs, traffic signals etc. Scene analysis,Computer-Aided Diagnosis (CAD), object recognition, image retrieval and image annotation are some of the applications that rely on the performance of a classification- based model.

Technical Details of Final Deliverable

1. Implementation of Scheme/Model and Results
2. Results Formulation Report and Final Thesis Submission 
3. Research Paper Submission report and acceptance result 

Final Deliverable of the Project Software SystemType of Industry IT Technologies Artificial Intelligence(AI)Sustainable Development Goals Quality EducationRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	80000
Laptop	Equipment	1	70000	70000
Software Licence	Miscellaneous	1	10000	10000