Scribd
Upload
15 views20 pages

Mini Project

The document outlines a project focused on developing a deep learning model for violence detection in images and videos, emphasizing its importance for safety applications. It discusses the methodology, tools used, and the challenges faced, including the need for contextual understanding and dataset diversity. The project aims to enhance real-time monitoring capabilities and suggests future improvements such as hybrid models and multi-modal detection.

Uploaded by

Pratik Anand
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
15 views20 pages

Mini Project

The document outlines a project focused on developing a deep learning model for violence detection in images and videos, emphasizing its importance for safety applications. It discusses the methodology, tools used, and the challenges faced, including the need for contextual understanding and dataset diversity. The project aims to enhance real-time monitoring capabilities and suggests future improvements such as hybrid models and multi-modal detection.

Uploaded by

Pratik Anand
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PPTX, PDF, TXT or read online on Scribd
You are on page 1/ 20

DEPARTMENT OF COMPUTER SCIENCE

AND ENGINEERING
RVS COLLEGE OF ENGINEERING AND
TECHNOLOGY

VIOLENCE DETECTION USING DEEP


LEARNING

PRATIK ANAND
CSE/169/23
PRAVEEN HEMBROM
CSE/170/23
PRAVEEN KUMAR MISHRA
CSE/171/23
PRIYANSHU KUMAR
UNDER THE GUIDANCE OF
CSE/172/23
PROF.KUMAR
PRIYANSHU NUSHRAT PRAVEEN
CSE/173/23
CONTENT
 Introduction
 Literature review
 Objective
 Research Gap
 Tools and technology used
 Pipeline
 Workflow overview
 Optimizer(adam vs adamax)
 Optimizer comparision
 Sample output
 Website page
 Result and analysis
 Dataset overview
 Future scope and improvement
 Conclusion
 references
INTRODUCTION
• Violence detection models are a key component in the intersection of
computer vision and safety applications, designed to identify and classify
violent or harmful content within images and videos .

• With the growing need for monitoring and filtering digital content, this
technology is increasingly being applied in areas such as public safety,
content moderation, and law enforcement .

• The primary goal of a violence detection model is to automatically assess


images or video footage and categorize them as either violent or non-violent

o This automation aids in monitoring vast amounts of content quickly, ensuring


timely intervention and preventing harm, whether in physical environments or
online platforms.
Literature
Author(s)
Review Methods/Algorithm
Key Contribution
Used
Year

Arun Akash S The study utilizes the Inception-V3 202


A, Sri Inception-v3 and Yolo-v5 1
Skandha models to detect violent
Moorthy R acts.
Sri Skandha The number of persons Yolo-v5 202
Moorthy R involved, and the 2
weapons used in real-
time from CCTV
footage.
Muhammad Violence detection using 3D CNN 201
Razman, spatiotemporal Feature 9
Adnan abid
Abbasi- Introduced a method to 3D ConvNet, SSIM 202
Moghadam extract significant 2
temporal information by
utilizing the SSIM
OBJECTIVE
• Our project focuses on building a model that can
automatically detect violence in images and videos.

• This is an important step towards creating safer


environments, whether it's online content
moderation, surveillance systems, or assisting law
enforcement in analyzing images for signs of
violence..

• We aim to expand the model's abilities in the future,


but for now, it serves as a foundational tool for
detecting violence in visual data.
Research Gaps
 Current models often struggle to understand the context
in which actions occur. For example, certain actions might
appear violent in isolation but are benign in the correct
context (e.g., a play fight).

 Many datasets used for training models may lack diversity


in terms of geographic, cultural, and environmental
contexts.

 Real-time violence detection at scale, such as in public


surveillance systems, remains challenging due to
computational constraints
Tools and Technology Used:

 Convolutional Neural Networks (CNNs): For image detection


 TensorFlow and Keras: For building and training models.
 Tf.data API: for preprocessing and dataset handling.
 Python Libraries : Matplotlib, NumPy for data handling and
visualising.
 Web Application : CSS&HTML, javascript.
 Fast api: Model deployment.
Pipeline of CNN Architecture
WORKFLOW OVERVIEW
DATASE Preprocessing and data Dataset
T Augmentation splitting

Web Model
Model architectu
applicatio deploy
training re
n.
STEPS DESCRIPTION
DATA COLLECTION Collected labeled datasets from Kaggle with
'Violence' and 'Non-Violence' categories.

DATA Resized images to 256x256 pixels,


normalized pixel values, and applied data
PREPROCESSING augmentation (rotation, flipping, cropping)
to enhance dataset diversity.
DATA SPLITTING Split data into Training (80%), Validation
(10%), and Testing (10%) sets to ensure
proper model training and evaluation

MODEL TRANING Used a CNN model to learn features from


the preprocessed images, focusing on
distinguishing violent and non-violent
content.
DEPLOYMENT Integrated the trained model into a web
application for real-time image classification,
enabling users to detect violence in uploaded
Optimizer(Adam vs Adamax)
 During training, the Adam optimizer was initially used, but
it resulted in the lowest validation accuracy. To improve
performance, the Adamax optimizer was implemented.

 Adamax not only provided higher validation accuracy but


also converged faster, reaching peak accuracy in fewer
epochs than Adam.

 This suggests that Adamax handles sparse updates better


and improves model generalization, making it the better
choice for this particular violence detection task.
Optimizer Comparision
Adam
Adamax
Results Analysis
Sample Output
Test image from
Raw image
dataset
Website page
Dataset Overview
NON-VIOLENCE
1. Source of Dataset
The images were collected from Kaggle,
a popular platform for datasets and
competitions.
2. Dataset Composition
The dataset is organized into
subdirectories, each representing
different categories:
Violence: Images depicting violent
scenes. VIOLENCE
Non-Violence: Images without any
violent content.
Total number of images: 11,1007 .
Violence: 5,231 images.
Non-Violence: 5,832 images.
3. Dataset Split
The dataset was divided into three
parts:
Training Set: 80%
Validation Set: 10%
FUTURE SCOPE AND
IMPROVEMENT
 Hybrid CNN-RNN Approaches:
• Instead of only using CNNs, combine CNNs with LSTMs
or GRUs to analyze both spatial (image) and temporal
(sequence) features

 Multi-modal Detection:
• Combining audio and video for more accurate detection.

 Dataset & Bias Reduction:


• Violence datasets are often limited to certain environments
(e.g., street fights, not domestic violence).
CONCLUSION
 Violence detection using machine learning and Python
can significantly enhance safety and security.

 Despite challenges, it has immense potential in real-time


applications such as surveillance and social media.

 Continuous research and advancements in AI will make


these systems more efficient and ethical.
REFERENCES
 [1] A. M. Ali, S. S. S. R. Depuru, and R. V. D. D. Kumar, Proceedings of
the 2021 IEEE International Conference on Data Science and
Advanced Analytics (DSAA), 2021.
 [2] A. Gupta, D. S. Chowdhury, and R. Mahapatra, Journal of Computer
Vision and Image Processing, vol. 6, no. 2, pp. 112-125, 2020.
 [3] R. N. Jena, S. P. Maji, and D. L. Dey, Proceedings of the IEEE
International Conference on Image Processing (ICIP), 2020.
 [4] A. Khan, T. N. Nguyen, and R. K. B. R. Reddy, Journal of Visual
Communication and Image Representation, vol. 60, pp. 121-132, 2019.
 [5] P. H. Nguyen, H. M. Le, and Q. V. H. Le, Proceedings of the 2018
International Conference on Image Processing and Vision Engineering
(IPVE), 2018.

You might also like