SRM INSTITUTE OF SCIENCE AND TECHNOLOGY

Ramapuram, Chennai – 600 089

SCHOOL OF COMPUTER SCIENCE AND ENGINEERING

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING

21CSP302L – PROJECT
End to end text processing for dynamic classification and high
precision keyword extraction
Batch No : 18 III year / VI Semester
Class / Sec : CSE - G
Team Members Supervisor

SANTHOSH KUMAR D[RA2211003020421] Mrs.L.Madhuri Devi

THARUN P[RA2211003020430] Assistant Professor / Department of Computer
NAVEEN M[RA2211003020385] Science
SRM Institute of Science and Technology
Ramapuram, Chennai

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 1

 AGENDA
• Abstract
• Introduction
• References (Base paper to be included)
• INTRODUCTION
• OBJECTIVES
• LITERATURE SURVEY
• PROBLEM STATEMENT
• EXISITING WORKS
• DISADVANTAGES
• PROPOSED SYSTEM
• ADVANTAGES
• CONCLUSION
• REFERENCES

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 2

 ABSTRACT
• Automated Text Classification: Dynamically classify text into categories using

state-of-the-art models like BERT.High-Precision Keyword Extraction: Accurately

extract relevant keywords/phrases for deeper content understanding.Real-time

Processing: Efficiently handle large text data in real-time, adapting to evolving

content.

• The goal of this project is to build an end-to-end text processing system that can

automatically classify text into relevant categories and extract high-precision

keywords or key phrases. By leveraging advanced machine learning models like

BERT, the system dynamically categorizes text based on its content, ensuring it

adapts to evolving information.

• .
D
3
EPARTMENT OF COMPUTER SCIENCE AND ENGINEERING
 ABSTRACT
• Additionally, the system extracts the most meaningful keywords and phrases using

a combination of traditional and deep learning techniques. This approach ensures

accurate and efficient text analysis in real-time, even for large datasets.

• The system’s dynamic classification component uses fine-tuned machine learning

models like BERT to ensure accurate categorization of text. These models are

designed to continuously adapt to new content and categories as the system learns

over time. For keyword extraction, the system combines traditional methods such

as TF-IDF and TextRank with deep learning to accurately identify the most

contextually relevant keywords and key phrases. This combination allows for

precise extraction of meaningful insights from text, improving content

understanding and analysis.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 4
 INTRODUCTION
• In today’s data-driven world, processing and analyzing vast amounts of

unstructured text has become a critical task across industries. Whether it's social

media posts, customer reviews, or research papers, extracting useful information

from text quickly and accurately is key to making informed decisions. This project

focuses on developing an end-to-end text processing system designed to classify

text dynamically and extract high-precision keywords or key phrases.

• By utilizing state-of-the-art machine learning models, such as BERT, for

classification and combining traditional and deep learning methods for keyword

extraction, the system can effectively categorize text into relevant categories while

identifying the most important terms that represent the essence of the content.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 5

 INTRODUCTION
• The system is built to process large volumes of data in real-time, adapt to new data

patterns, and provide valuable insights for a wide range of applications, from

content categorization to enhancing search and recommendation systems..

• The overall goal of this project is to provide a scalable, real-time solution that can

handle large datasets, analyze text at scale, and deliver precise, actionable insights.

Whether it’s improving content recommendations, enhancing search capabilities, or

helping businesses understand customer sentiment, this system will empower users

to process and analyze text in a faster, smarter, and more meaningful way.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 6

 OBJECTIVES
• The system will clean and organize raw text, removing irrelevant content and
normalizing the data.Using models like BERT, the system will classify text into
relevant categories and adapt to new content.
• It will also extract high-precision keywords by combining traditional methods like
TF-IDF with deep learning for accuracy and context.The system will process large
volumes of text in real-time, ideal for applications like social media monitoring and
customer feedback analysis.
• An API will be developed for easy integration with other applications.Performance
will be monitored for accuracy and scalability to handle growing data. Ultimately,
the system will provide actionable insights, enabling better decision-making from
accurate content analysis.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 7

 LITERATURE SURVEY
[1] Yuan et al. discusses image style transfer, a technology that merges the style of one
image with the content of another to create new artworks. It highlights the role of deep
learning, particularly CNNs like VGG-19, in automating this process. Traditional
methods relied on manual techniques but were limited in scalability. The introduction
of neural style transfer (NST) by Gatys et al. (2016) marked a significant advancement,
enabling the separation and recombination of content and style features. Subsequent
developments include real-time style transfer, multi-style techniques, and arbitrary
style transfer methods like AdaIN.

[2] Longqing et al. delves into the application of convolutional neural networks
(CNNs), particularly the VGG-19 model, for image style transfer, a technique that
merges the artistic style of one image with the content of another to create visually
compelling artworks.
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 8
 LITERATURE SURVEY
The process involves preprocessing and postprocessing functions, with VGG-19
extracting features from both content and style images. A weighted loss function,
combining content loss, style loss, and total variation loss, is minimized to balance
content preservation, style transfer, and noise reduction. The style is represented using
Gram matrices, ensuring the synthesized image retains the original content while
adopting the style's color palette. Introduced by Gatys et al. in 2015, this method has
gained significant traction in computer vision and image processing, with VGG-19
emerging as a preferred model due to its robust feature extraction capabilities. Recent
advancements, such as those by Liao et al. and Wu et al., have further expanded its
applications, including in areas like rice disease classification.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 9

 LITERATURE SURVEY
[3] Rohit et al. presents a hybrid deep learning model for generating captions from
images, combining computer vision and natural language processing. The model uses
VGG-19, a convolutional neural network, to extract feature vectors from images and an
LSTM network to generate descriptive captions. Trained on datasets like FLICKS and
SK, the model's performance is evaluated using the BLEU score. Image captioning,
which involves understanding and describing image elements and their relationships,
has applications in aiding visually impaired individuals, enhancing image searches, and
improving online marketing. Previous approaches include methods by Farhadi et al.
and Kulkarni et al., which focus on scoring image-sentence relevance and generating
natural language descriptions, respectively. The proposed model integrates VGG-19
and LSTM to address the complexity of caption generation, offering a robust solution.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 10

 LITERATURE SURVEY
[4] Prathamesh et al. proposes an innovative approach to Artistic Style Transfer (AST)
using the VGG-19 neural network to address the challenge of color distortion in Neural
Style Transfer (NST). Traditional NST methods often alter the color palette of the
content image, compromising its authenticity. The proposed method integrates VGG-19
for extracting style features while employing luminance transfer techniques to preserve
the original color harmony of the content image. This ensures that the final output
retains the natural coloration of the content image while incorporating the artistic style
of the chosen painting. The approach was tested on various content images,
successfully maintaining color fidelity throughout the style transfer process. By
combining VGG-19's robust feature extraction capabilities with luminance transfer, this
research enhances the quality and authenticity of digitally transformed images, offering
a valuable tool for artists and technologists.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 11

 LITERATURE SURVEY
[5] Yilin et al. applies deep learning-based convolutional neural networks (CNNs) for
image style transfer, highlighting the advantages of using pre-trained models like
VGG-16 and VGG-19. Traditional methods of image style transfer, which relied on
non-photorealistic rendering techniques, were surpassed by the introduction of neural
style transfer (NST) by Gatys et al. in 2015. NST leverages CNNs to extract and fuse
content and style features from images, producing results that retain the semantic
content of the original image while adopting the artistic style of another. The study
compares the performance of VGG-16 and VGG-19 for this task, concluding that
VGG-16 offers better efficiency and quality in style transfer. CNNs, with their layered
structure (input, convolutional, activation, pooling, and fully connected layers), excel
at extracting both shallow and deep image features, making them ideal for style
transfer.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 12

 LITERATURE SURVEY
[6] Yakun et al. explores the concept of image style transfer, a technique that converts
an image into a different artistic style while preserving its original content. This
process has become increasingly popular in image processing applications, driven by
advancements in deep learning and convolutional neural networks (CNNs). The paper
reviews various methods of neural style transfer, starting with the seminal work by
Gatys et al., which uses CNNs to extract and statistically represent content and style
features from images. This approach involves iteratively optimizing a white noise
image to match the content and style of target images, using pre-trained models like
VGG-16. While effective, this method is computationally intensive, requiring hundreds
of iterations.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 13

 LITERATURE SURVEY
[7] Jyoti et al. proposes a system aimed at assisting visually impaired individuals by
generating detailed scene descriptions using large language models (LLMs) like GPT2,
DistilGPT2, BERT, and RoBERTa. The system utilizes an Encoder-Decoder
architecture, with Vision Transformers (ViT) as the encoder and a distilled GPT-2
model as the decoder, to create comprehensive image captions. Trained on a diverse
dataset of around a hundred thousand samples using advanced hardware, the model
achieved a ROUGE score of 21.69%, demonstrating its potential to produce human-
like descriptions. This research has significant implications for enhancing the
independence, education, and employment opportunities of visually impaired
individuals by improving their awareness of surroundings.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 14

 LITERATURE SURVEY
[8] Vidyadevi et al. explores advancements in image captioning, a task that combines
image feature extraction using Convolutional Neural Networks (CNNs) and natural
language generation via Long Short-Term Memory (LSTM) networks to create
descriptive captions for images. The paper surveys recent methods, highlighting
models like Inception-v4 CNN for encoding images and LSTM for generating
sentences, achieving a BLEU-1 score of 0.758367 on the Flickr 8k dataset.
Applications include aiding visually impaired individuals, enhancing publishing
workflows, and improving medical diagnostics. The literature review covers techniques
such as adversarial caption generation, which refines captions for relevance, and the
use of datasets like MS COCO and Situitertock Images for training. The proposed
CNN-LSTM model demonstrates strong performance, with future enhancements
focusing on improving accessibility.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 15

 LITERATURE SURVEY
[9] Suganthe et al. worked on an advanced approach to image captioning, which
involves describing images based on their features and actions. Traditional methods use
an encoder-decoder structure with CNNs for feature extraction and LSTMs for caption
generation, often facing issues like gradient explosion and inefficient information
extraction. To address these challenges, the proposed model combines YOLOv5 for
object detection and Bidirectional LSTM (Bi-LSTM) for feature extraction and caption
generation. YOLOv5 divides images into grids to detect objects efficiently, while Bi-
LSTM processes the extracted features to generate descriptive captions. This approach,
tested on the Flickr8k dataset, outperforms traditional methods by leveraging local
image features rather than relying solely on global features. The model's performance
is evaluated using the BLEU score, achieving a score of 0.7, indicating its effectiveness
in generating accurate captions.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 16

 LITERATURE SURVEY
[10] J.K.Periyasamy et al. propses the use of deep learning models for the early
detection of brain tumors from MRI scans, a critical task given the high mortality rate
associated with brain tumors. Traditional methods of tumor detection are time-
consuming and prone to human error, prompting the need for automated, computer-
assisted diagnosis systems. The study compares the performance of two deep learning
models, VGG-19 and ResNet50, in detecting brain tumors using MRI images. These
models leverage image processing and deep learning techniques to analyze MRI data,
reduce noise, and improve accuracy in tumor detection. The research highlights the
advantages of using CNNs, which are widely employed in image classification tasks,
over traditional machine learning approaches, despite the latter being faster. The goal is
to provide pathology specialists with effective tools for accurate and timely diagnosis,
ultimately improving patient outcomes.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 17

 PROBLEM IDENTIFICATION
• Overwhelming Volume of Unstructured Text: With the increasing amount of
unstructured text data from sources like social media, customer feedback, blogs,
and articles, businesses and organizations are facing difficulties in efficiently
processing and analyzing this data. The sheer volume makes it nearly impossible to
manually categorize and extract insights without the risk of human error and bias.
• Inefficient Traditional Methods: Current methods for text classification and
keyword extraction often rely on outdated algorithms or simplistic approaches like
keyword matching, which fail to account for context, semantics, or evolving
language patterns. Traditional methods like TF-IDF, while useful, don't always
provide the accuracy needed for real-time applications or adapt well to new topics
and trends..

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 18

 PROBLEM IDENTIFICATION
• Lack of Scalability and Adaptability: These systems are often unable to handle
large datasets efficiently, limiting their ability to provide timely insights.
Additionally, they may not be able to dynamically adjust to new categories or
evolving language, reducing their effectiveness as data patterns shift over time.
• Limited Real-Time Processing: Real-time applications, such as social media
monitoring or customer sentiment analysis, require systems that can process text
quickly and accurately.
• Need for Actionable Insights: Without the ability to automatically and accurately
classify text and extract meaningful keywords, businesses and researchers struggle
to gain actionable insights from their data. This hinders their ability to make
informed, data-driven decisions and to identify key trends or emerging topics in a
timely manner..
DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 19
 EXISTING SYSTEM
• There are several existing systems that aim to classify text and extract keywords,
but many face limitations when it comes to handling large, dynamic datasets.
Traditional methods, like rule-based systems and Bag of Words (BoW), can be
simple to implement, but they often struggle to understand context and adapt to
new language or evolving topics.
• While machine learning models like Naive Bayes and Support Vector Machines
(SVM) can help classify text, they often require retraining when new topics arise
and may not perform well with more complex or nuanced text. Features like real-
time customization, caching of precomputed style weights, or descriptive
captioning for enhanced usability are entirely absent, which reduces user appeal.
• More advanced systems, like Latent Dirichlet Allocation (LDA) and deep learning
models such as BERT, show great promise for understanding context and
improving accuracy DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 20
 DISADVANTAGES
• High Computational Needs: The system requires significant computational
resources, especially for real-time processing with advanced models like BERT,
which could increase costs.
• Complex Training: Training deep learning models demands large labeled datasets
and expertise, which can be time-consuming and resource-intensive.
• Model Interpretability: Deep learning models like BERT are complex and hard
to interpret, which can be a drawback when users need transparency in decision-
making.
• Scalability Issues: While the system is designed to scale, processing extremely
large datasets in real-time may still challenge the infrastructure.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 21

 PROPOSED SYSTEM
• The proposed system will leverage advanced machine learning and natural
language processing techniques to automate text classification and keyword
extraction. It will begin by preprocessing raw text to remove irrelevant content and
standardize it, ensuring it’s ready for analysis.

• The system will then classify the text into relevant categories using BERT, a state-
of-the-art deep learning model that adapts to new content and evolving language. In
addition to classification, the system will extract high-precision keywords and key
phrases by combining deep learning models like BERT with traditional methods
such as TF-IDF, ensuring that the extracted terms are both contextually accurate
and meaningful.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 22

 PROPOSED SYSTEM
• The system is designed to handle real-time data, enabling it to process large
volumes of text quickly and efficiently. This makes it suitable for use cases like
social media monitoring, customer feedback analysis, and real-time trend detection.
Furthermore, a user-friendly API will be developed to allow seamless integration
with other applications, enabling users to easily send text data and receive
classification results and extracted keywords. The system will also be scalable,
ensuring it can grow with increasing data volumes and continue to deliver high-
quality results over time, making it an ideal solution for organizations looking to
extract actionable insights from large text datasets.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 23

 ADVANTAGES
• High Accuracy and Precision: By combining advanced deep learning models like
BERT with traditional methods like TF-IDF, the system ensures that both text
classification and keyword extraction are highly accurate. BERT’s contextual
understanding of text allows for better interpretation of complex language, leading
to more precise classifications and relevant keyword extractions compared to
traditional systems.
• Real-Time Processing: The system is designed to handle large volumes of text in
real-time. This capability is crucial for applications where timely insights are
needed, such as social media monitoring, sentiment analysis, and customer
feedback analysis. Real-time processing enables businesses to quickly react to
emerging trends or issues, providing a competitive edge.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 24

 ADVANTAGES
• Scalability: The system is built to scale, allowing it to handle increasing amounts
of text data without sacrificing performance. As organizations generate more data
over time, the system will be able to process and analyze larger datasets efficiently,
ensuring that it can grow alongside the business needs.
• Flexibility and Adaptability: With the use of BERT, the system can adapt to new
and evolving content. Unlike traditional rule-based systems, which require manual
updates to handle new topics, the BERT model continuously learns and evolves,
making the system highly flexible. This ensures that it can classify text and extract
keywords from emerging trends and topics without requiring extensive retraining.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 25

 CONCLUSION
• The proposed system successfully combines artistic style transfer and image

captioning into a unified platform, offering users a creative and functional

experience. By utilizing powerful deep learning models like VGG-19 and ResNet-

50, the system ensures high-quality style transformation with support for both

predefined and custom styles.

• The incorporation of precomputed style weights enhances efficiency, making the

system suitable for real-time applications without compromising on output quality.

Additionally, the comparative analysis feature provides unique insights into the

strengths of different architectures, giving users greater control and understanding

of the results.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 26

 CONCLUSION
• This system goes beyond conventional style transfer by integrating image

captioning, adding an accessibility layer that caters to a broader audience, including

visually impaired users. Its modular and scalable design enables future

enhancements, such as expanding the style database or introducing new features.

• The project serves as a robust solution for blending creativity, technology, and

accessibility, with potential applications in art, education, and multimedia

industries. By addressing the limitations of existing systems, the proposed system

paves the way for innovative advancements in the field of artistic transformation

and AI-driven creativity.

DEPARTMENT OF COMPUTER SCIENCE AND ENGINEERING 27

 REFERENCES
[1] https://ieeexplore.ieee.org/document/9731923 - Image Style Transfer with Feature
Extraction Algorithm using Deep Learning (tick)
[2] https://ieeexplore.ieee.org/document/10405398 - New Image Processing: VGG
Image Style Transfer with Gram Matrix Style Features
[3] https://ieeexplore.ieee.org/document/9579897 - Hybrid Feature and Sequence
Extractor based Deep Learning Model for Image Caption Generation
[4] https://ieeexplore.ieee.org/document/10593426 - Optimizing Color Preservation in
Artistic Style Transfer with VGG-19 and Luminance Transfer Approach
[5] https://ieeexplore.ieee.org/document/9918891 - Image Style Transfer Based on
VGG Neural Network Model
[6] https://ieeexplore.ieee.org/document/10075728 - Image Style Transfer–A Critical
Review
[7] https://ieeexplore.ieee.org/document/10704129 - Comparative study of different
LLM's for Captioning Images to Help Blind People

DEPARTMENT OF COMPUTER SCIENCE AND 28

ENGINEERING
 REFERENCES
[8] https://ieeexplore.ieee.org/document/9726074 - Image Caption Generation using
Deep Neural Networks
[9] https://ieeexplore.ieee.org/document/10393602 - Leveraging Deep Learning Model
for Image Caption Generation for Scenes Description
[10] https://ieeexplore.ieee.org/document/10128287 - Image Captioning Using Deep
Learning Techniques for Partially Impaired People
[12] https://ieeexplore.ieee.org/document/10126451 - Comparison of VGG-19 and
RESNET-50 Algorithms in Brain Tumor Detection

DEPARTMENT OF COMPUTER SCIENCE AND 29

ENGINEERING