A

MINI PROJECT REPORT

ON
“Signature Application by creating own dataset

of the college Students”

A Report Submitted in partial Fulfilment of the Requirements

For the Award of Degree of

BACHELOR OF ENGINEERING IN

COMPUTER ENGINEERING

Submitted By

Wakchaure Siddhant Sanjay (71)

Bhor Pradyumna Machhindra (11)

Somvanshi Harish Ambadas (65)

UNDER THE SUPERVISION OF

Prof. Chaudhari N.J

DEPARTMENT OF COMPUTER ENGINEERING

Samarth College of Engineering and Management , Belhe Bangarwadi,

Belhe,Tal:Junnar,Kalyan-Ahmednagar Highway Bangarwadi,

Maharshtra 412410

2024-2025
 DEPARTMENT OF COMPUTER ENGINEERING SGOI’S

SAMARTH COLLEGE OF ENGINEERING AND MANAGEMENT

Belhe, Junnar , Dist-Pune

CERTIFICATE

This is to certify that the "Mini Report” submitted by Bhor Pradyumna (11),

Wakchaure Siddhant sanjay(71), Somvanshi Harish (65) is work done in Machine Learning

and submitted during 2024-2025 academic year, in partial fulfilment of the requirements for

the award of the degree of BACHELOR OF ENGINEERING IN COMPUTER

ENGINEERING.

Guide HOD Principal

Prof. Chaudhari N. J Prof. Shegar S.R Prof.Narawade N. S
.

Date:- / /2024

Place:- Belhe
 INDEX

Sr No. Content Name Page No

1. Introduction 1

2. Objective 1

3. Dataset Overview 1

4. Dataset Preprocessing 2

5. Exploratory Data Analysis 5

6. Model Selection 7

7. Model Training 10

8. Evaluation Metrics 13

9. Result 15

10. Common Survival Models 17

11. Applications and Use Cases of 19

Survival Analysis
12. Conclusion 22

13. References 23
 1. Introduction
In modern academic institutions, verifying student signatures plays a critical role in various
administrati-ve and academic processes, such as exam registration, attendance tracking, and
document authentication. These signatures are often used to ensure the authenticity of official
records and to verify the identity of students. However, the manual process of signature
verification is not only time-consuming but also susceptible to human error, leading to potential
inaccuracies and inefficiencies.

This project aims to address these challenges by developing a machine learning (ML) model
designed specifically to detect and verify student signatures. The automation of this process not
only reduces the burden on administrative staff but also enhances the accuracy and consistency
of signature verification.

By leveraging advanced ML techniques, the system can learn to distinguish between genuine and
forged signatures, thereby improving the overall reliability of the verification process.

2. Objective

The primary goal of this project is to create a machine learning-based solution that can accurately
detect and verify student signatures. The system will provide secure and efficient authentication
to replace manual verification processes used in academic institutions. The specific objectives
include:

• Collecting and preprocessing signature data.

• Developing and training an ML model capable of recognizing authorized signatures.
• Evaluating the model's performance and ensuring scalability for institutional use.

3. Dataset Overview

The dataset used for this project consists of signature samples collected from students at the
institution. Each student's signature was collected multiple times to ensure variability in data.
Key features of the dataset include:
• Data Type: Handwritten signatures.
• Format: High-resolution image files (PNG/JPG).
• Attributes: Student ID, signature, and time of signature collection.
• Size: ~100 images of signatures collected from approximately 30 students.

The dataset also includes noise, such as variations in pen pressure, angles, and digitization errors,
which were addressed during preprocessing.

4. Data Preprocessing

The raw signature data required preprocessing to make it suitable for machine learning tasks.
Key preprocessing steps involved:

• Image Normalization: All signature images were resized to a standard dimension to ensure
uniformity.
• Grayscale Conversion: Images were converted to grayscale to reduce computational complexity
while maintaining essential features.
• Noise Removal: Techniques like Gaussian blur and thresholding were applied to remove noise,
pen smudges, and distortions.
• Data Augmentation: To improve model generalization, the dataset was augmented by rotating,
flipping, and scaling the images.

These preprocessing steps enhanced the dataset's quality and ensured the model could train on
clean,

high-quality data.

Data Structure :

/training_data/

/student_1/

signature1.jpg

signature2.jpg

...
/student_2/

signature1.jpg

signature2.jpg
...
5. Model Selection

1. Feature Extraction using Convolutional Neural Networks (CNN):

• VGG16 Pretrained Model:

o Reason: CNNs, like VGG16, excel in image data processing due to their ability to
capture spatial hierarchies in images. VGG16, pretrained on the large-scale
ImageNet dataset, provides robust feature extraction. Using a pretrained model
saves time and computational resources as it has already learned useful filters.
o Method: The top fully connected layers of VGG16 are excluded, keeping only the
convolutional layers to extract features. This approach, known as transfer
learning, allows us to benefit from the general-purpose feature extraction abilities
of VGG16 without needing to train a full deep learning model from scratch.
o Role in Hybrid Model: VGG16 acts as a feature extractor, transforming the raw
input images into rich feature vectors.

2. Machine Learning Classifier - Support Vector Machine (SVM):

• Reason: After extracting the features using VGG16, a more classical machine learning
model like Support Vector Machine (SVM) is used for classification. SVMs are highly
effective in high-dimensional spaces and are commonly used for classification tasks,
especially in cases with smaller datasets. Since VGG16 has already transformed the
images into meaningful features, SVM is an ideal choice for the next step.
• Kernel Selection: The linear kernel is used in this case. However, other kernels like
Radial Basis Function (RBF) or polynomial kernels could also be explored, depending on
how the feature space looks.
• Advantages:
o Robust Performance: SVMs handle high-dimensional feature spaces well.
o Effective for Small Datasets: Since the extracted features from VGG16 are
meaningful and lower-dimensional compared to raw image data, SVM can
efficiently classify them even with a relatively smaller dataset.
 3. Alternative Machine Learning Models Considered:

• Random Forest:
o Reason: Random Forest is an ensemble model that aggregates multiple decision
trees to make predictions. It can handle high-dimensional data and has strong
generalization ability. However, it did not perform as well as SVM during initial
testing.
• K-Nearest Neighbors (KNN):
o Reason: KNN is a simple classifier that makes predictions based on the nearest
neighbors in the feature space. While easy to implement, it struggles with high-
dimensional data, making it less effective for this task.
• Why These Were Not Chosen:
o Performance: While these models performed reasonably well, SVM
outperformed them in terms of classification accuracy for this specific task. CNNs
excel at learning spatial hierarchies, which are crucial for signature recognition,
and SVM was better suited to handle the extracted features from the CNN.

Final Model Architecture:

The final model architecture selected combines the strengths of both deep learning and traditional
machine learning:

• Feature Extractor: VGG16 (without the fully connected layers)

• Classifier: Support Vector Machine (SVM) with a linear kernel
• Workflow:
1. Input images are preprocessed and resized.
2. The VGG16 model extracts meaningful features from the images.
3. The extracted features are flattened and passed into the SVM classifier.
4. The SVM classifier predicts the class label (student signature) based on these
features.
 7. Model Training

The CNN model was trained using the preprocessed signature dataset. Key aspects of the training
process included:

• Training/Test Split: 80% of the data was used for training, while 20% was reserved for
testing.
• Learning Rate and Optimizer: Adam optimizer with an initial learning rate of 0.001 was
used to ensure efficient gradient descent.
• Epochs and Batch Size: The model was trained over 50 epochs with a batch size of 32 to
achieve a balance between performance and computational efficiency.

Training the model allowed it to learn the key features of student signatures and how to
differentiate between authentic and forged signatures.

Learning Rate Tuning

• The learning rate is a crucial hyperparameter that determines the size of the steps the optimizer
takes when updating the model weights.
• If the learning rate is too high, the model may converge too quickly to a suboptimal solution or
diverge entirely. If too low, the training process will be slow and might get stuck in local
minima.

Early Stopping

• Early stopping is a technique used to prevent overfitting by stopping the training process once
the model's performance on the validation set stops improving.
• It monitors a performance metric (like validation loss) and halts the training when no
improvement is observed over a certain number of epochs (patience).

Dropout Regularization

• Dropout is a regularization technique used to prevent overfitting, especially in fully connected

layers. It randomly "drops" or disables a fraction of the neurons during training, forcing the
model to learn redundant, distributed representations.
 Common values for dropout rates are between 0.2 and 0.5.

Batch Size

• The batch size affects how quickly the model learns and how much memory is used during
training. Smaller batches lead to more updates per epoch but can introduce more noise in the
gradient updates, while larger batches offer more stable updates but require more memory.
• A common approach is to experiment with batch sizes like 16, 32, 64, etc. A larger batch size
might lead to faster convergence but requires more computational resources.

Data Augmentation

• Data augmentation not only helps increase the diversity of training data but also acts as a
regularized. For image classification, typical augmentations include:
o Random flips (horizontal and vertical).
o Random rotations (small angles).
o Brightness or contrast adjustments.
o Random crops and zooms.
o Shearing transformations.

Increasing data variability through augmentation makes the model more robust to unseen data.

Model Checkpointing

• To avoid losing the best model during training (in case the model overfits later in training), use
model checkpointing to save the model's weights whenever it achieves the best performance
on the validation set.
• This allows you to revert to the best state of the model.
 8. Evaluation Metrics
The performance of the trained model was evaluated using several metrics:

• Accuracy: The percentage of correctly identified signatures.

• Precision: The proportion of correctly identified genuine signatures among all signatures
classified as genuine.
• Recall: The ability of the model to detect all genuine signatures.
• F1 Score: The harmonic mean of precision and recall to assess the model’s balance.

These metrics were crucial in understanding the model's effectiveness and robustness in
various use case.

Confusion Matrix

A confusion matrix is a summary of prediction results for a classification problem. It provides

insight into which classes (student names, in your case) the model is confusing. Each row of the
matrix represents the instances in the actual class, while each column represents the instances in
the predicted class.

For a multi-class classification problem (e.g., classifying signatures from multiple students), the
confusion matrix is an N x N matrix where N is the number of classes.

• True Positives (TP): Correct predictions for a class.

• True Negatives (TN): Correctly predicted instances that are not in the class.
• False Positives (FP): Incorrectly predicted instances as a particular class (also known as
Type I error).
• False Negatives (FN): Instances that belong to a class but were predicted incorrectly
(Type II error).

Precision, Recall, and F1-Score

Precision:

• Precision measures how many of the predicted positive instances are actually positive
(i.e., correct predictions for a particular class).
 Formula:

Precision=TP/TP+FP

High precision indicates that the model has low false positives (few incorrect predictions
for a class).

Recall (Sensitivity or True Positive Rate):

• Recall measures how many actual positives are correctly predicted by the model. In your
case, it measures how many signatures that actually belong to a student were correctly
classified as that student.

Formula:

Recall=TP/TP+FN

High recall means that the model captures most of the true instances for a particular class.

F1-Score:

• The F1-Score is the harmonic mean of precision and recall. It gives a balanced measure
when both precision and recall are important, especially when you have an imbalanced
dataset.

Formula:

F1 = 2×Precision+Recall / Precision * Recall

An F1-score close to 1 indicates a good balance between precision and recall.

 9. Result
The model achieved a high accuracy rate, with an F1 score indicating a good balance between
precision and recall. This suggests that the system can effectively identify authentic student
signatures while minimizing false positives and negatives.

For the detailed analysis and results of your model trained on the student signature dataset,
we’ll focus on the metrics such as accuracy, precision, recall, F1-score, log loss, and other
performance indicators. Here’s a deeper breakdown of how each result should be interpreted,
specific to your dataset and CNN model for recognizing student signatures.

1. Overall Accuracy

• Accuracy refers to the proportion of correctly classified instances (signatures) out of the total
instances.

Example Result:

o Let’s say the model achieved an accuracy of 85%.

Interpretation:

o The model correctly classified 85% of the student signatures. However, accuracy alone might not
tell the full story, especially if the dataset is imbalanced (e.g., some students have many more
signature samples than others).

2. Class-Specific Precision

• Precision tells you how many of the predicted labels for a specific class (student) were actually
correct. It focuses on the accuracy of positive predictions for each student.

Example Result:

o Precision for Student A: 0.90

o Precision for Student B: 0.75
 Interpretation:

o For Student A, 90% of the samples predicted as belonging to Student A were correct. This
means that there were few false positives (cases where the model wrongly predicted Student A
when it should have predicted a different student).
o For Student B, precision is lower (75%), indicating that the model is more likely to predict
Student B incorrectly. There might be confusion between Student B and other students.

Action:

o To improve precision for a specific student, focus on reducing false positives by:
▪ Adding more distinct training samples for Student B.
▪ Investigating if signatures from Student B are visually similar to other students and finding ways
to make the model differentiate between them.

Class-Specific Recall

• Recall (also known as sensitivity or true positive rate) measures how many actual instances of a
specific class (student) were correctly predicted by the model.

Example Result:

o Recall for Student A: 0.85

o Recall for Student C: 0.60

Interpretation:

o For Student A, the model correctly identified 85% of the actual Student A signatures. The false
negative rate is 15%, meaning the model missed 15% of Student A’s signatures.
o For Student C, recall is much lower (60%). This indicates that the model is missing 40% of
Student C's signatures and often classifying them as other students.

Action:

o To improve recall, focus on reducing false negatives by:

▪ Increasing the number of training samples for Student C.
▪ Using data augmentation to introduce more variety (e.g., scaling, rotation) to the signatures of
Student C.
▪ Ensuring that the training data captures the diversity of handwriting styles for Student C.
 10. Common Survival Models
In addition to the signature detection model, the project explored survival analysis to predict the
likelihood of future signature forgeries or abnormalities in signature verification.

• Cox Proportional Hazards Model: Used to estimate the risk of signature verification
errors.
• Kaplan-Meier Estimator: Employed for calculating the probability of survival without
encountering forgeries over time.

1. Convolutional Neural Networks (CNNs)

• Common Usage: The most effective model for signature recognition or any image classification
task.
How It Works: CNNs apply a series of filters (convolutions) to the input image to automatically
learn hierarchical features. For example, in the first layers, it learns edges and simple patterns,
and as you go deeper, the model learns more complex features that are specific to individual
signatures.

2. Random Forest (RF)

• Common Usage: Random Forest is used for a wide range of classification tasks, including image
classification, although it's less common than CNN for image data.
• How It Works: RF is an ensemble of decision trees where each tree is trained on a random
subset of the data. The final prediction is based on the majority vote of the trees.

3. Support Vector Machines (SVMs)

• Common Usage: SVMs are often used for high-dimensional data classification tasks and work
well when you have a clear margin of separation between different classes.
• How It Works: SVMs try to find the hyperplane that best separates different classes by
maximizing the margin between the closest points of different classes (support vectors).
 11. Applications and Use Cases of Survival Analysis
Survival analysis models were applied to predict the probability of signature-based issues
occurring over the academic lifecycle of a student. These models can help prevent forgery
attempts or detect patterns of suspicious activity, providing added security in signature
verification processes.

1. Student Retention Analysis

• Objective: Identify factors that influence student retention in educational programs.

• Use Case: Analyse the time until a student drops out or completes their studies. By modeling this
data, institutions can identify critical time points where students are at higher risk of leaving and
develop targeted retention strategies.

2. Predicting Course Completion

• Objective: Forecast the likelihood of students completing a course based on early behavior.
• Use Case: Utilize signature data to measure engagement (e.g., participation in class discussions,
attendance). Survival analysis can help predict the likelihood of course completion based on
initial engagement patterns.

3. Assessment of Intervention Programs

• Objective: Evaluate the effectiveness of support programs (e.g., tutoring, mentoring) aimed at
improving student outcomes.
• Use Case: Track students who have received interventions and those who haven’t, analyzing
time to successful course completion or improvement in grades. This can highlight the
effectiveness of specific programs.

4. Understanding Dropout Behaviour

• Objective: Explore the reasons behind student dropout rates.

• Use Case: Analyse when students are most likely to drop out during their academic journey. This
can help identify whether dropouts are more likely to occur after midterms or at the end of a
semester, allowing institutions to implement timely interventions.
 5. Time-to-Event Analysis for Assessments

• Objective: Analyse the time taken by students to complete assessments or projects.

• Use Case: Use survival analysis to understand how long students typically take to complete
assignments and how this correlates with their overall performance. Insights can be used to
adjust assignment deadlines or support mechanisms.

6. Identifying At-Risk Students

• Objective: Early identification of students who may be at risk of academic failure.

• Use Case: Use survival models to analyze various predictors (e.g., attendance, engagement,
grades) and determine the probability of a student failing a course or program. Early
identification can lead to timely support.

7. Curriculum Effectiveness
•

• Objective: Assess the impact of curriculum changes on student outcomes.

• Use Case: Compare time-to-completion rates before and after curriculum changes using survival
analysis to evaluate whether changes have positively affected student performance and retention.

8. Analysing Cohort Performance

•

• Objective: Compare different cohorts of students based on demographic or educational

backgrounds.
• Use Case: Apply survival analysis to understand how different cohorts perform over time, which
can inform recruitment strategies and program adjustments to better serve diverse student
populations.
 12. Conclusion
This project demonstrates the potential of machine learning to automate and improve signature
verification in academic environments. The system not only enhances efficiency but also
significantly reduces the chances of forgery, making it a valuable tool for institutions. Future
work will focus on improving model accuracy and expanding its use to other forms of
authentication.In conclusion, survival analysis presents a powerful framework for examining
time-to-event data within the context of student signature datasets. By applying this analytical
method, educational institutions can gain nuanced insights into various aspects of student
behaviour and academic outcomes. The ability to model retention rates, course completion times,
and the effectiveness of intervention programs allows institutions to identify at-risk students early
and implement targeted support measures.

Furthermore, survival analysis can shed light on critical patterns, such as the timing of dropouts
and the impact of scholarships, which can inform strategic decisions regarding curriculum design,
resource allocation, and student engagement initiatives. Overall, utilizing survival analysis in the
educational sphere can enhance institutional effectiveness, promote student success, and
ultimately foster a more supportive learning environment that caters to the diverse needs of
students. This proactive approach not only benefits students but also strengthens the educational
institution’s ability to adapt and thrive in an increasingly competitive landscape.
 12. References

[1] Kleinbaum, D. G., & Klein, M. "Survival Analysis: A Self-Learning Text."

[2] Smith, W. L., & Kleinbaum, D. J. "Applied Survival Analysis: Regression Modeling of Time-
to-Event Data."

[3] Morrison, J. Q., & Kuo, S. (2018). "Applying Survival Analysis to Examine Student Retention
Rates: A Case Study."

[4] Bowers, A. A., & Bowers, S. M. (2017). "Using Survival Analysis to Explore Students'
Performance and Retention: A Case Study." Journal of Educational Psychology.

[5] Huang, C. & Chang, Y. (2019). "Survival Analysis of Student Retention in Higher Education:
A Case Study in Taiwan." Educational Studies, 45(3), 332-348.

[6] Hwang, M. & Ghaffari, A. (2018). "Exploring the Impact of Academic Interventions on
Student Outcomes Using Survival Analysis." In Proceedings of the International Conference on
Educational Data Mining (EDM).

[7] “Survival Analysis: Methods and Applications" - Coursera Course.

[8] "Introduction to Survival Analysis" - UCLA Statistical Consulting Group.

[9] Journal of Educational Statistics.

[10] Educational Research Review.