Image Processing and Classification

A Project Report by
Anjali Tripathi
Roll No.2200190210008

Under the Guidance of

Ms. Swati Mishra
Assistant Professor, Dept. of EEE

2023-24
JSS MAHAVIDYAPEETHA
JSS ACADEMY OF TECHNICAL EDUCATION, NOIDA
DEPARTMENT OF ELECTRICAL AND ELECTRONICS ENGINEERING
C-20/1, SECTOR-62, NOIDA, UTTAR PRADESH-201301

0
 PAGE NO.

CONTENTS 1

LIST OF FIGURES 2

LIST OF TABLES 3

Chapter-1: Introduction 4-5

Chapter-2: Objectives and Scope of the study 6

Chapter-3: Theoretical Background 7-11

Chapter-4: Results 12-13

Chapter-5: Conclusion 14

Chapter-6: Limitations 15

Chapter-7: Future Work 16

References 17

1
 LIST OF FIGURES

Figure No. Description Page No.

Figure 1 Image Classification 4

Figure 2 CNN Architecture 8

Figure 3 Different Classes of Skin Cancer 9

Figure 4 Distribution of data in Different classes 10

of Ham10000

Figure 5 Architecture of MobilenetV2 11

Figure 6 Model Accuracy 12

Figure 7 Model Loss 12

Figure 8 Classification Report 13

2
 LIST OF TABLES

Table No. Description Page No.

Table-I Different classes of Ham10000 dataset 10

Table-II Classification Report 13

3
 CHAPTER-1
INTRODUCTION

Image Processing and Classification

Image processing and classification are fundamental tasks in
computer vision that play a crucial role in various fields such as
medical imaging, surveillance, autonomous vehicles, and more. This
report aims to provide an overview of image processing techniques
and classification methods in detection of skin cancer.

Skin diseases represent a significant health concern globally,

affecting millions of individuals.Image processing and classification
techniques play a vital role in the early and accurate diagnosis of
various skin conditions.This report provides an overview of image
processing methodologies and classification algorithms used in the
detection and classification of skin diseases.

In our mini project we have used HAM10000 and ISIC Challenge

2016 large-scale dataset, where HAM10000 contains large number
of skin cancer disease and ISIC Challenge 2016 contain large data
about skin cancer disease and about patients.

figure1: Image Classification

4
In this report we explore the application of CNNs for image
processing and classification using the HAM10000 and ISIC
Challenge 2019 datasets. By leveraging deep learning, we aim to
create accurate models (by using Resnet50, MobileNetV2 and
many more) for automatic skin lesion detection and classification.

Further , we have run the model such as Mobile NetV2 and

Resnet50 and obtained the various parameters such as accuracy,F1
Score,Recall value and precision etc.

5
 CHAPTER-2
OBJECTIVES AND SCOPE OF STUDY
OBJECTIVE:

The primary objectives of this study are

1. To explore and analyze image processing techniques for
preprocessing skin lesion images.
2. To execute CNN models for classification using the HAM10000
and ISIC Challenge 2019 datasets.
3. To evaluate the performance of the trained models in classifying
skin lesions.

SCOPE OF STUDY:

● DATASET UNDERSTANDING

● IMAGE PRE-PROCESSING

● SEGMENTATION TECHNIQUES

● FEATURE EXTRACTION

● CLASSIFICATION USING CNN

6
 CHAPTER-3
THEORETICAL BACKGROUND

IMAGE PROCESSING

Image processing and classification techniques in skin disease

analysis involve the utilization of computational methods to enhance,
analyze, and classify images of skin lesions. This interdisciplinary
field at the intersection of computer science and dermatology plays
a crucial role in the early detection, diagnosis, and management of
various skin conditions.

Here are some techniques involved in image Processing…

● Image Enhancement
Adjusts the contrast of an image by redistributing pixel intensities.
Increases the dynamic range of pixel values to enhance contrast.
Enhances image details by emphasizing edges and high-frequency
components. Filters out unwanted noise while preserving important
image features.

● Image Restoration
Removes blur caused by motion, defocus, or other factors to restore
image sharpness. Reduces noise in an image using techniques like
median filtering, Gaussian filtering, or wavelet denoising.Inpainting: Fills
in missing or damaged regions of an image based on surrounding
information.

CNN
A Convolutional Neural Network (CNN) is a type of Deep Learning neural
network architecture commonly used in Computer Vision. Computer vision is a
field of Artificial Intelligence that enables a computer to understand and
interpret the image or visual data.
figure 2:CNN Architecture

7
 Figure 2: Image Classification
The Architecture of a convolutional network typically consists of four
types of layers:
• Convolution
• Pooling
• Activation
• Fully Connected

1. Convolutional Layers: These layers apply learnable filters

(kernels) to input images, convolving them to produce feature maps.
Filters capture local patterns and spatial dependencies within the
input data, enabling the network to detect meaningful features at
different scales and orientations.

2. Pooling Layers: Pooling layers down sample feature maps by

aggregating information within local regions. Common pooling
operations include max pooling and average pooling, which reduce
computational complexity,increase translation invariance, and
promote spatial hierarchies in feature representation.

3. ReLU Activation Layers: The convolution maps are passed

through a nonlinear activation layer, such as the Rectified Linear
Unit (ReLu), which replaces negative numbers of the filtered images
with zeros.
8
4. Fully Connected Layers: These layers integrate features
extracted by preceding layers to perform high-level reasoning and
classification. They connect every neuron to every neuron in the
subsequent layer, facilitating the learning of complex, nonlinear
relationships between features and target labels.

HAM10000
The HAM10000 dataset comprises a diverse collection of
dermatoscopic images depicting various types of pigmented skin
lesions. With a total of 10,000 images distributed across seven
distinct classes, this dataset serves as a valuable resource for
training and evaluating skin lesion classification models.

The Dataset contains seven skin lesion classes:

● Melanoma.
● Neveus.
● Basal Cell Carcinoma.
● Actinic Keratosis.
● Benign Keratosis.
● Dermatofibroma.
● Vascular Lesion.

Figure 3: Different classes of Skin Cancer

9
 Figure 4:Distribution of data in different classes of HAM10000 Dataset:

SKIN LESIONS SYMPTOMS Cancerous/Non-Cancerous

MELANOMA Bleeding,Itching or Crushing Cancerous

NEVES Pain,Bleeding or Inflammation Non-Cancerous

BKL A round or oval-shaped,waxy or Non-Cancerous

rough bump

BCC Shiny,pink or flesh-colored papule Non-Cancerous

AKIEC Color variations,Red or Brown Non-Cancerous

VASC Flat pink areas of skin discoloration Non-Cancerous

DF Flesh-colored,brown or Non-Cancerous
reddish-brown

Table 1:Different Classes of HAM10000 Dataset

10
Training Model used in Ham10000
Model Used:MobileNetV2

MobileNet-v2 is a convolutional neural network that is 53 layers

deep. You can load a pretrained version of the network trained on
more than a million images from the ImageNet database [1]. The
pretrained network can classify images into 1000 object
categories, such as keyboard, mouse, pencil, and many animals.

It is a powerful and lightweight model for image classification tasks.

Its efficient architecture, combined with its ability to maintain high
accuracy, makes it an ideal choice for resource-constrained devices.

We got accuracy as 74% on running Mobilenetv2 model.

figure 5:CNN Architecture of MobileNetV2

11
 CHAPTER-4
RESULTS

● Results of Mobilenetv2 on running for 25 epochs on the given dataset:

figure 6:Model Accuracy

Figure is showing the accuracy with respect to epoch.
Here is the highest accuracy at epoch 21th that is 79.84%.

figure7:Model Loss
Figure is showing the model loss with respect to epoch.
Here we get the minimum loss at 25th epoch which is 0.5469

12
 PERFORMANCE EVALUATION METRICS

Table 2 :Classification Report

13
 CHAPTER-5
CONCLUSION

• Till now we have read the HAM10000 Dataset in detail and implemented
the CSV data reading and image reading algorithms.

• Also, we have gone through the basics of feature extraction and the basics
of the CNN model and its performance evaluation metrics.

• We have implemented algorithms to train our model to get the desired

output on the HAM10000 Dataset, although the accuracy of output was
not satisfactorily acceptable.

14
 CHAPTER-6
LIMITATIONS

• Variability in performance due to factors like dataset quality, model

architecture, and hyperparameter settings.

• Limited computational resources and training data may hinder model

generalization.

• ISIC dataset(or any other large dataset) requires heavy models and a
high-end systems to run the algorithms.

15
 CHAPTER-7

FUTURE WORK

• Exploration of advanced CNN architectures: Experiment with

different CNN architectures such as ResNet, VGG, or EfficientNet to
improve model performance and robustness.

• Transfer learning techniques: Utilize pre-trained CNN models on

larger datasets like ImageNet and fine-tune them on skin lesion data
to leverage learned features and enhance classification accuracy.

• Improving the accuracy of the HAM10000 Dataset.

• To implement the algorithms on the ISIC Challenge 2019 Dataset.

• Integration of clinical metadata: Incorporate additional clinical

metadata such as patient demographics, medical history, and lesion
characteristics into the classification models to improve personalized
diagnosis.

16
 References
[1] Codella, N. C. F. et al. (2019). Skin Lesion Analysis Toward Melanoma Detection: A
Challenge at the 2017 International Symposium on Biomedical Imaging (ISBI), Hosted by
the International Skin Imaging Collaboration (ISIC). arXiv preprint arXiv:1902.03368.
[2] Tschandl, P. et al. (2018). The HAM10000 dataset, a large collection of multi-source
dermatoscopic images of common pigmented skin lesions. Scientific Data, 5(1), 180161.
[3] Dataset Link: https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000
[4] Code used: https://github.com/anitadala
[5] Science direct research paper:
https://www.sciencedirect.com/science/article/pii/S1877050923016277

17