Adichunchanagiri Institute of Technology

Chikmagalur -577102
Department of Computer Science and Engineering
Internship Project seminar on
“ Natural Scene Recognition”

Carried at

Submitted By:

Varun A.S(4AI16CS117) Sourab Saklecha(4AI16CS094)

Samrudh Patel(4AI16CS075) Sandesh S (4AI16CS077)
 CONTENTS
• Abstract
• Introduction
• Motivation
• Problem Statement
• Objectives
• Literature Survey

• System Requirement Specification

• System Design
• Results and Discussion
• Conclusion and Future enhancement
• References
 ABSTRACT

 Missing, disappearing, swapping are fundamental problems encountered

especially in pet animals that are very similar.

 Traditionally, animals are known through their external appearances and patterns.

 Biometric based systems developed for the identification of animals are scarce and
many of these systems are inadequate

 To simplify the process, we propose an intelligent pet animal recognition system,

which is developed by using the CNN model which is used to classify the pet
animals into different groups.

 The recognition process of animals will be faster and intelligent using the pet
animal recognition system by reducing human involvement.
 INTRODUCTION
 Despite of advances in software and technology, nowadays, detection and
recognition of scenes is a process and there is no single method that provides an
effective solution

 The problem in the detection and identification of natural scenes is approached in

the form of double pattern recognition.

 The goal is to classify images under 6 categories. {'buildings' -> 0, 'forest' -> 1,
'glacier' -> 2, 'mountain' -> 3, 'sea' -> 4, 'street' -> 5 } by using a CNN model and
a DenseNet121 and ResNet50 model (pretrained on imagenet) for the
classification tasks.

 Thus, a system which recognizes the individual scenes based on their biometric 
features required in order to help in  identification and classification.
This is a sample dataset which is divided into 6 different scenes of nature
All the pictures of the images have been resized down to 512 x 384.

Figure 1: Images of different Natural Scenes.

 MOTIVATION

 To develop an automated system which is able to identify the types of

nature scenes.
 Reconstruction of 3D scenes from images .
 High dimensional image spaces.
 Automatic analysis of images .
PROBLEM STATEMENT
To design and implement a model for the Natural Scene recognition using machine
learning technique. The main aim of this project is to detect the category of the
scene.

Input: Images of different categories of nature scenes.

Processing: • Train the CNN with at least 3 convolution layers

• Data Augmentation procedure with accuracy reaching at least 99%
and also take care of over fitting.

Output: The trained CNN should able to recognize the natural scenes and classify
it as one of the 6 categories.(example output: given image is of buildings)
Fig 2 : Basic System Architecture of Image Classification
OBJECTIVES
The Objective of the project are:

 This project gives a general idea of how image classification can be done efficiently.
 The scope of the project can be extended to the various industries where there is a
huge scope for automation by just alerting the data set which is relevant to the
problems.
 Aim is to make the model learnt to distinguishing features of natural scenes.
 To analyze the image of nature by using shape and color.
 It is not possible reference, our model focuses on capturing natural scene based on its
feature learnt.
SCOPE
The Scope of the project are:

 It can be used in pattern, object, and event recognition.

 The analysis of the shape and color geometry features set helps in the
recognition of natural scenes based on the features.

 The system machine learning capabilities and the image processing framework
helps in the recognition of natural scenes.
 LITERATURE SURVEY

SL Author & Paper Description Publication

No. Title

Xiaoming Huang, Text detection and Conference on estima-

Tao Shen , Run recognition in natural scene -tion, Detection and
1. Wang , Chenqiang images plays an important Information Fusion
Gao. “Text detection role in content analysis of (ICEDIF) 2015
and recognition in images. International
natural scene images

N. Ezaki , M. Bulacu We propose a system that Proceedings of the 17th

, L. Schomaker. reads the text encountered in International
2. “Natural scene natural scenes with the aim to Conference on Pattern
detection: towards a provide assistance to the Recognition, 2004.
system for visually visually impaired persons. ICPR 2004.
impaired person
SL No. Author and Paper Title Description Publication

Y. Chai, V. Lempitsky, The processing on sematic ICCV, 2011.

3.
and A. Zisserman. and visual information has
Bicos: A bi-level co- always been a challenging
segmentation method for task and co-segmentation is
image classification one of them. Co –
segmentation is a classic.

P. Silapachote , J. Visually impaired IEEE Computer

4.
Weinman , A. Hanson , individuals are unable to Society Conference
M.A. Mattar , R. Weiss. utilize the significant on Computer Vision
“Automatic Sign amount of information in and Pattern
Detection and signs. Recognition 2005.
Recognition in Natural
Scenes.

Table 1: Literature Survey for Natural scene recognition

SYSTEM REQUIREMENT SPECIFICATION

Operating system : Windows 7/ Windows 10

Coding Language : Python

Tool : Google Colaboratory

 SYSTEM DESIGN
Learning
Algorithm

Image
Dataset
Train
Train Set Model

Image
Feature Validate
Pre-
Extraction Model
processing

Test Set
Predicted
Result

Fig 3: System Architecture of the proposed method

USE CASE DIAGRAM Data collection

Data preprocessing

Feature analysis

Feature selection

Data transformation

Data preparation for

training

Model construction

USER Model training

Validate Model

Result analysis

Fig 4: An Use-Case Diagram of natural scene classification

 Source Task Classification

Datasets CNN

Model Model Model Transferring learnt

learnt weights for new Target
from
Task

Learnable Weights

Knowledge

Fig 5: Flow Chart of proposed system

 IMPLEMENTATION

The implementation includes following modules :

 Image Input

 Implementing CNN Algorithm

RESULTS AND DISCUSSION

Fig 6 :Snapshot of Forest scene recognition.

Fig 7 : Snapshot of Sea scene recognition.
Fig 8 :Snapshot of Building recognition.
Fig 9 :Snapshot of Street recognition.
Fig 10 :Snapshot of Glacier recognition.
Fig 11: Snapshot of Mountain recognition.
 CONCLUSION AND FUTURE ENHANCEMENT

 The Natural scene recognition is possible through machine learning and computer
vision algorithms. To create a more accurate system, there must be a large and
growing data set.
 This system recognizes the type of the scene using the picture in a cost effective
way.
 The system can be upgraded to the better detection of waste by using advanced
algorithms of machine learning.
 This system further can be made advanced to recognize the scenes during day,
night and different weather conditions like fog, rain, etc.  
 REFERENCES
[1]. O.M. Parkhi, A, Vedaldi, A. Zisserman, C.V. Jawahar. “Cats and Dogs”, IEEE Conference on
Computer Vision and Pattern, 2016.

[2]. Derek Hoiem, Yodsawalai Chodpathuwan, and Qieyun Dai. “Diagnosing Error in Object
Detectors”. In proc. ECCV,2017.

[3].Y. Chai, V, Lempitsky, and A. Zisserman. Bicos: A bi-level co-segmentation method for
image classification. In proc, 2018.

[4].W.Zhang, J. Sun, and X. Tang. “Scene detection-how to effectively exploit shape and tex
THANK YOU