Query optimization and processing for advanced database systems
URBAN OBJECT DETECTION IN UAV RESNETpptx
1. DEVELOPING THE RESNET
NEURAL NETWORK FOR
URBAN OBJECT DETECTION IN
UAV
ACHARYA NAGARJUNA UNIVERSITY
BY :
M . LEELA PHANI L21EC3269
G . PARDAV Y20EC3214
P . YADIDYA Y20EC3242
UNDER THE GUIDANCE OF :
PROF.P.SIDDAIAH .
M.TECH,PH.D .
3. ABSTRACT
In this project, we have proposed a cutting-
edge model for object classification of UAV
(unmanned aerial vehicle) captured urban
area images dataset. We have used COCO-
dataset and PASCAL VOC which is a popular
dataset for aerial images classification and
detection using deep network designer
present in MATLAB.
4. WHAT SHOULD YOU KNOW FIRST
UAV's object detection plays a crucial role in
fields such as agriculture, disaster response,
surveillance, and environmental monitoring.
The integration of CNN in UAV enables to
autonomously detect and classify objects in
real world
CNNs are particularly effective for this task
due to their ability to learn hierarchical
features from images
5. S.NO TITLE OF
SAMPLE PAPER
YEAR AUTHO
R
TECHNIQUE ACCURACY LIMITATIONS &
DRAWBACKS
1 Computer vision
methods for 2D object
detection from UAV
2020 Marco Leo RGB based
Object detection
83.4% • Only for indoor
navigation and
applicable for
only limited hieght
2 Object Detection in
UAV images via
Global density Fused
convolution Network
2020 Xiao
Haung
GDF net (global
density fused
convolution
Network)
85.26% • Requires more
computational
power
• Many layers in the
network increases
time
LITERATURE REVIEW
3 COMNET:
Combinationa neural
network for object
detection in UAV
borne thermal images
2020 Jiasong Li Combinational
Network
87.4% • It is only
applicable for
night vison and
processes only
thermal images
only
6. S.
NO
TITLE OF SAMPLE
PAPER
YEAR AUTHOR TECHNIQUE ACCURACY LIMITATIONS &
DRAWBACKS
4 Multiscale object
Detection from drone
imagery using ensemble
transfer learning
2021 Abali
malathe
Ensemble learning
ResNet
94.56% • It can’t detect new
objects
• Can’t able to
handle complex
scenes
5 Object Detection from the
video taken by drone via
convolution neural
Network
2020 Chefan sun Tensor flow object
detection API
96.54% • Computationally
more Expensive
6 On board small sacle
object detection for
Unmanned Aerial Vehicle
(UAV)
2023 Muhamma
d yousaf
YOLO, SSD(single
shot multibox
Detector)
91.26% • Small target sizes
• Low resolution
occlusion
• Data Set scarcity
7. S. NO TITLE OF SAMPLE
PAPER
YEAR AUTHOR TECHNIQUE ACCURACY LIMITATIONS &
DRAWBACKS
7 Evaluating the influence
of backbone of network
architectures for object
detection in aerial
images
2023 Kanhg
nquyen
Faster R-CNN
,backbone
networks
77.64% • Only focuses on
vehicle detection
• Only used to two
data sets which
may not represents
the real world
aerial images
8 Adaptive dense pyramid
network for object
detection in UAV
imagery
2022 Yufeng wang VGG-16
Backbone
SSD
84.26% • Network has higher
computational
complexity
• It does not consider
the rotation and
occlusion of objects
8. The journals confined that existing object
detection methods based on Convolutional
Neural Networks (CNNs) often suffer from
poor generalization and robustness when
applied to UAV images, as they are usually
pre-trained on natural scene images that have
different characteristics and distributions.
Therefore, there is a need for an effective and
efficient system that can detect urban objects
from UAV images using a deep learning model
that is specifically designed and trained for
UAV data.
PROBLEM STATEMENT
9. In this project, the utilization of ResNet is
pivotal to enhance the efficiency of
object detection
Res Net architectures are versatile and
applicable to various CNN tasks
including object classification
YOLO (you only look once) is neural
network architecture is best efficient
detector for real time object detection
PROPOSED METHOD
11. ResNet, short for Residual Network, is a deep
learning architecture known for its ability to
train very deep neural networks effectively.
Its skip connections allow for easier
optimization and can be beneficial for object
detection tasks in UAVs.
The ResNet architecture introduces the
simple concept of adding an intermediate
input to the output of a series of convolution
blocks
Res Net
13. TRAINING PLOT
• The Accuracy of ResNet 101 pre trained Convolution Neural Network with 0.01
learning rate using Sgdm Optimizer is nearly 95% with a certain data set
14. IN [ ] : clc;
net = trainedNetwork_3;
data = image Datastore ("D:project
filesdatalearningtest","IncludeSubfolders",true,"LabelSource","foldernames")
;
in = augmentedImageDatastore ([224 224],data);
Labels = data. Labels;
predict = classify(net, in);
accuracy = (sum(Labels == predict, "all“)/numel(predict))*100;
disp(accuracy);
confusion chart(Labels, predict)
stats = statsOfMeasure(confusion mat(Labels, predict),1);
%Process a sequence of files.
MATLAB SYNTAX
15. IN [ ] : imds = image Datastore ("D:project filesdata learningvalidate", "include
Subfolders",true,"Label Source“,"Folder names");
% Create an image Datastore
%Get all filenames into one cell array. Filenames have the complete path
(folder prepended).
all Filenames = imds. Files;
Num Files = numel(imds. Files);
for k = 1 : numel(all Filenames)
% Get this file name.
full Filename = all Filenames{k};
in_data = imresize(imread(full Filename),[224 224]);
[Label, Probability] = classify(net, in_data);
figure;
16. IN [ ] : imshow(in_data);
title({char(Label),(max(Probability))})
end
OUT [ ] :
19. FUTURE WORK :
Using the trained network we classify the images of
urban area captured by satellite and try to improve the
accuracy of object classification compared with
traditional method of object classification which has less
accuracy in object classification.