Volume 7, Issue 11, November – 2022 International Journal of Innovative Science and Research Technology

ISSN No:-2456-2165

Animal Intrusion Detection System using

CNN and Image Processing
K Bhumika1, Sirisha Madhuri T.2, G. Radhika3, CH Ellaji4
M. Tech1, M. Tech (Ph.D)2,3,4
Assistant professor (Senior)2, Assistant professor 3,4
Sri Padmavathi mahila university (Tirupati)

Abstract:- One of the greatest dangers to agricultural (ANN) to help farmers overcome their ordeal and improve
productivity is animal damage to agriculture. Crop accuracy in identifying plant diseases.
raiding has become one of the most antagonistic human-
wildlife conflicts as cultivated land has expanded into On the other hand, the precision of such systems is
previous wildlife habitat. Farmers in India endures major heavily reliant on feature selection methodologies. With the
risks from pests, natural disasters, and animal damage, all elimination of image pre-processing and the facilitation of
of which result in lesser yields. Traditional farming intrinsic feature selection, convolutional neural networks
methods are unsuccessful and hiring guards to watch have recently made significant progress in automated image
crops and keep animals at bay is not a practical solution. recognition. The availability of a large enough dataset for
It is critical to protect crops from animal damage while such competitions is another problem. In situations when
also redirecting the animal without injuring it, as the there is a lack of data, it is best to utilise a model that has
safety of both animals and people is essential. To get over already been trained on a big dataset. Transfer learning is used
these obstacles and accomplish our goal, we employ the to develop a model that is then fine-tuned to adapt the final
deep learning concept of convolutional neural networks, a layers to make them unique to the dataset at reference.
subfield of computer vision, to identify animals as they
enter our farm. The primary goal of this project is to However, the precision of such systems is significantly
constantly monitor the entire farm using a camera that influenced by feature selection procedures. Convolutional
records the surroundings at all hours of the day. We neural networks (CNNs) are a kind of neural network that
significantly improves image-based recognition by doing
identify animal infiltration using a CNN algorithm
and Xgboost and notify farmers when this occurs. away with the need for picture pre-processing and allowing
intrinsic feature selection. Another difficulty is acquiring a
Keywords:- CNN, XGBoost, Computer vision. large dataset sufficient for these contests. When the dataset
size is limited, it is better to use a model that has been pre-
I. INTRODUCTION trained on a large dataset. This is known as Transfer Learning.
The dataset was developed to help with animal spotting using
A. GENERAL INTRODUCTION information gathered over the course of many months.
In India, agriculture plays a key role. In India, agriculture Various animal images are used for the detection of the
has always been a major contributor to the country's GDP. intrusion in farming fields, for the training the model.
The majority of Indians find work in agriculture, yet this
sector is filled with challenges. Animal-human conflict has  CONVOLUTIONAL NEURAL
long been a big problem that has wasted countless resources  NETWORK
and put people's lives in danger. In recent years, there has One kind of ANN is the convolutional neural network
been an increase in the frequency of these sorts of conflicts. (CNN). Since it is optimised for handling data in pixel
Therefore, this area needs continual surveillance to deter the format, it may be used for image recognition and
introduction of such animals and other undesirables. processing. It has multiple layers to process the data
Encroachment and poaching lead to human-animal conflicts and extract important features from the grid like
because people go into the forest for subsistence, claiming arrangement.
land for agricultural practises and fast industrialisation leads
the growth of urban ground, while animals approach CNN saves a lot of time and also reduces errors.
surrounding villages in search of water during the dry Since, it doesn’t need more parameters for learning the
summer months. Elephants and wild boars often trample characteristics of image filters. The main purpose of
agricultural plants in their hunt for food. When one species the CNN algorithm is to extract image features without
needs something, it might threaten the survival of another, losing the data it represents. CNN is based on
depleting resources and sometimes claiming lives. South Asia neuroscience findings and is made up of artificial
and Africa are the most common regions where people and neurons.
elephants come into conflict with one another. In most cases,
an electric fence is used to keep animals out of a farm's crops. Usually for image processing through CNN, if an
The address situations the animals to act in an odd manner, image is given then it takes pixel values as data and
making them easier to capture. Deep learning strategies are picks some visual features. It points out the pixel
required to avoid these types of issues. Researchers used values of colors for the activation function.
various machine learning strategies similar to the Support
Vector Machine (SVM) and Artificial Neural Networks

IJISRT22NOV797 www.ijisrt.com 197

Volume 7, Issue 11, November – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
At initial stage it takes the edges of the picture III. PROPOSED MODEL
and sends it to detecting the corners and color groups
of the image. Then the image defined is passed to the The proposed system consists of algorithms which are
next layer for pooling, this cycle continues until the used in detecting animal intrusion in human lands and alerting
image is predicted. the farmer when the intrusion occurs. Based on the CNN
classifier, here the database for animals are collected and
The decreased size of the image gives finite trained for predicting the intrusion. Two stage CNN with
definition which makes RoI pooling occur. It only binary and multiclass level are used for the classification and
reveals the most relevant features for activation feature extraction. Here, make use of LBP comprehension
function features for dense process. and XGBoost count, which provides the join mastermind by
a large area of interest comprising both target animals and
other false positive targets. A two stage flow is proposed for
the high rate results regardless of less time and enormous
usage. RoI pooling is considered for the multiple object
detection in a single image.

The entire process can be divided into four modules

such as:
 Training images
 Data augmentation
 CNN classification
 Prediction and Alert

A. TRAINING IMAGES:
The training images is done to obtain the images of
animals and train the own database set. By this the animals
which intruded into farming fields are detected through
Fig. 1: CNN model cameras located on the fields. Here various animals which are
B. OBJECTIVES common for the field intrusion are considered such as deer,
The goal of the project is to give security against wild cows, wild boar, elephants, monkeys etc. Kaggle is used for
animal attacks, ultimately decreasing the farmer's potential downloading image dataset for detection of animal intrusion.
loss. B. DATA AUGMENTATION
 Detecting intrusions in the field. Using a process called data augmentation, fresh training
 Capturing the intruder's image and identifying it using data may be produced from existing data. It helps prevent
image processing. machine learning models from being overfit during training
 Contrasting the image with an image of an animal. by acting as a regularizer. Here in animal intrusion detection
 Taking appropriate response based on the type of intrusion. large datasets are considered for training as such the data
augmentation leads to better prediction accuracy. While the
II. LITERATURE REVIEW process is operating, it has no effect on the size of the test
sample. Data augmentation uses the techniques such as
There is some related work that apply deep learning
position, scaling, cropping, padding, brightness etc., by these
methodology for detection of animal intrusion such as
the diversity of data within available is increased. Here data
Prakash, Banupriya[1] they describe a focused literature
augmentation is performed for the animal image data training
about the algorithm that classifies based on their image, so we
and prediction.
can effectively guide them. Animal detection and
classification may assist to avoid the accidents, track animals C. CNN CLASSIFICATION
and prevent theft. Nagaraju Andavarapu and Valli Kumari Convolutional neural network (CNN) classification is a
Vatsavayi[2] it states that WCoHOG is a Histogram oriented form of deep neural network used for image classification and
gradient based feature vector with improved accuracy. It is an computer vision applications. CNN's key benefit over its
expansion of Co-occurrence Histograms of Oriented forerunners is that it can recognise key traits automatically,
Gradients (CoHOG). LIBLINEAR classifier is used in order without human intervention. Convolutional neural networks
to get better accuracy for large dimensional data. (ConvNets) are both more effective and less asset than
traditional methods for machine learning. Here Two stage
S. Yadahalli, A. Parmar and A. Deshpande[3]Here
CNN is used such as binary level classifier and multi-level
various sensors and cameras are interfaced with the board
classifier. It is a supervised learning model which can
such as Passive Infrared Sensors(PIR), the camera capture the
differentiate various objects sizes, color in a single image.
intruding image with TFT display. Gyanendra K. Verma and
Return of Image pooling with XGboost is performed for the
Pragya Gupta [4] The multilayer graph cut in the spatial-
animal intrusion detection.
temporal domain was applied to a camera-trap database
including prospective animal suggestions. These are put to
use in the verification stage, which determines whether or not
a particular patch represents an animal.

IJISRT22NOV797 www.ijisrt.com 198

Volume 7, Issue 11, November – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
D. PREDICTION AND ALERT For the LBP, we first determine the LBP code for each
The number of the animals types to be predicted are of pixel and then build a histogram of those LBP codes to use as
approximately 20 with the image instances of 300 with the LBP feature. Assigning a value of 1 to a neighbouring
various angles for testing. The classified output saves the pixel x if and only if that pixel is also a neighbour of pixel p,
human time and alert the farmer of the intrusion and also the the central pixel's 8 neighbours are compared with pixel p.
cost damage caused to the field through intrusion of such
animal. The immediate alert will reduce the damage rate to C. XGBoost:
minimum. The prediction is performed by image testing XGBoost is an efficient and dependable gradient boosting
through deep learning classifiers. solution that stretches the boundaries of processing power for
boosted tree algorithms, with the primary goal of enhancing
the efficiency and effectiveness of machine learning models.
Steps involved in the processing of XGBoost are
Step 1: Load all libraries
Step 2: Load the dataset
Step 3: Data Cleaning and Feature Engineering
Step 4: Fine-tune and run the model
Step 5: Evaluate the Test Population

XGBoost's objective function is the product of a loss

function assessed over all predictions and a regularisation
function for all predictors (j trees). A prediction from the jth
tree, denoted by fj in the formula.
𝑛
𝑗
𝑜𝑏𝑗(𝜃) = ∑ 𝑙(𝑦𝑖 − 𝑦𝑖^ ) + ∑ 𝛺(𝑓𝑖 )
𝑗=1
𝑖
Fig. 2: Architecture model of proposed system
Log loss is a popular measure that is used by XGBoost.
IV. ALGORITHM Use this probability-based measure to evaluate a
classification model's efficacy.
A. Multipart CNN classifier:
A CNN has numerous layers. When these layers are 1
𝑁

utilised multiple times, they result in a Deep Neural − ∑ 𝑦𝑖 . log(𝑝(𝑦𝑖 )) + (1 − 𝑦𝑖 ). log(1 − 𝑝(𝑦𝑖 ))
𝑛
Network's path of action. There are three basic types of layers 𝑖=1

used to construct a CNN: Constructed trees and models are used to get all of the
 Information: This layer contains the image's incorrect potential solutions to the equation F2(x) =
pixel estimations. (0+1*h1(x)+1*h2(x)), where the final value of F2(x) is the
 Convolutional Layer: This layer receives the results of the XgBoost model prediction, if we are interested in creating a
neuron layer, which is linked to the data district. In this prediction for a new data point.
layer, we illustrate the degree of wreckage. Each channel
may be a 5x5 window that slides through the data and V. RESULT
selects the pixel with the highest striking power as the
yield. Our model has been trained by using Multistage CNN
 Modified Linear Unit [ReLU] Layer: This layer handles with LBP histograms and XGBoost gradient. Login page and
picture data in a region-wise manner. We like that a CNN registration page is created for the farmers for registering
employs back increase. Designers use the ReLU work to their details and when the animal is detected of intrusion in
maintain proportional estimations of the pixels while not the farming fields. The mail and alert notification is sent to
being affected by the back starting. the farmer. The Kaggle website is used for the image dataset
 Pooling Layer: This layer performs a down-testing work for detecting animal intrusion. The MySQL is used for the
along the spatial estimations (width, stature) in order to details storage of users and Tkinter is used as its frontend.
achieve volume.
 Totally Connected Layer: This layer is utilised to select
the score classes, i.e. which class has the best score when
compared to the data digits.

B. LBP:
To assign labels to picture pixels, Local Binary Pattern
(LBP) thresholds the area around each pixel and treats the
output as a binary number. To extract hierarchical data
representations, deep learning is widely used. The Local
Binary Pattern Network (LBPNet) is presented as a means of
efficiently collecting and comparing high-level over-
complete features in a multilayer hierarchy. Fig. 3: Home page

IJISRT22NOV797 www.ijisrt.com 199

Volume 7, Issue 11, November – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
The registration page and login page is of

Fig. 4: Registration page

Fig. 7: Graphical representation

VI. CONCLUSION AND FUTURE WORK

The approach performed responds to animal intrusion

on the farming fields. When the animal is detected upon
intrusion, then an alert notification is sent to the farmer
immediately. The animal captured is predicted through CNN.
LBP and XGboost are used for the feature extraction of the
image for the prediction. Through this crop damage can be
reduced. The dataset is collected from Kaggle. The
prediction obtained through this model is with better accuracy
and satiable result.

In the future, ViT and lightGBM will be used to achieve

faster and more accurate results. When compared to CNN,
Visual Transformers (ViT) performs better in terms of
Fig. 5: Login page computational efficiency and accuracy. and achieves better
results.
And result is shown as
REFERENCES

[1.] Parinita Badre, Siddhant Bandiwadekar, Prachi

Chandanshive, Aakanksha Chaudhari, Sonali Jadhav.
Automatically Identifying Animals Using Deep
Learning. In: 2018 International Journal Of Innovative
Research In Computer And Communication
Engineering (IJIRCCE). Vol. 6, Issue 3, March
(2018). Issn(online): 2320- 9801 ISSN (Print): 2320-
9798.
[2.] Gullal Singh Cheema, Saket Anand. Automatic
Detection And Recognition Of Individuals In
Patterned Species. IIIT – Delhi.
[3.] Automatically identifying, counting, and describing
wild animals in camera-trap images with deep
learning. In November (2017). arXiv:1703.05830v5.
[4.] Tibor Trnovszky, Patrik Kamencay, Richard Orjesek,
Fig. 6: Result Miroslav Benco, Peter Sykora. Animal Recognition
System Based on Convolution Neural Network. In
The graphical representation of Training loss and September (2017). DOI: 10.15598/aeee.v15i3.2202.
accuracy with the parameters of loss/accuracy and epoch [5.] Samer Hijazi, Rishi Kumar, Chris Rowen. Using
Convolution Neural Networks For Image
Recognition. IP Group, Cadence

IJISRT22NOV797 www.ijisrt.com 200

Volume 7, Issue 11, November – 2022 International Journal of Innovative Science and Research Technology
ISSN No:-2456-2165
[6.] Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun.
Faster R-CNN: Towards Real-time Object Detection
With Region Proposal Networks.
Arxiv:1506.01497v3 [Cs.CV] 6 Jan (2016).
[7.] HU W, HUANG Y, WEI L, ZHANG F, LI H (2015)
Deep Convolutional Neural Networks For
Hyperspectral Image Classification. Journal Of
Sensors 2015.
[8.] Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN:
Towards Real-time Object Detection With Region
Proposal Networks. In: NIPS. Pp. 91–99 (2015).
[9.] Simonyan, K., Zisserman, A.: Very Deep
Convolutional Networks For Large-scale Image
Recognition. Arxiv Preprint Arxiv:1409.1556 (2014).
[10.] Krizhevsky, A., I. SUTSKEVER And G. E. Hinton.
Imagenet Classification With Deep Convolutional
Neural Networks. Annual Conference On Neural
Information Processing Systems (NIPS). Harrah’s
Lake Tahoe: Curran Associates, 2012, Pp. 1097–1105.
Isbn 978- 1627480031.

IJISRT22NOV797 www.ijisrt.com 201