Acne Classification with Deep Learning

All fundamental DL4J dependencies are included in the pom.xml. The following project is built by me and my friend, Liang Han Sheng.

Introduction

Without a doubt, most of teenages face acne problem. However, there isn't a guideline on how serious the acne is and what proper steps should be taken in order to cure the acne and prevent scars. Hence, acne classification is developed using deep learning. It is carried out through transfer learning method using VGG-16 model in this project. It is able to classify acne seriousness into:

1. Normal
2. Level 0
3. Level 1
4. Level 2

and provide different solutions for user to cure the acne issue as shown below:

Values shown is the confidence level of different classes in percentage

Table of contents

1. Introduction
2. Contents
3. Dataset
4. Annotation
5. Data Preprocessing
6. Running on IDE
7. Running on Terminal
8. Results of AI Model
9. Future Improvements
10. Additional Information

Dataset

250 HD images were being hand-picked for each classes from various internet sources.

Annotation

Data annotation is being carried out by separating dataset into 4 classes.

Data Preprocessing

To increase the size of the dataset for training, data preprocessing is being carried out which includes:

• Horizontal Flip
• Vertical Flip
• 15° rotation
• 30° rotation

To run on IDE

1. Import project
2. Wait for IDE to resolve dependencies
3. Navigate to MyFirstDL4JProject.java
4. Run program

To run from command line

Firstly, the project needs to be compiled as a jar file. The command used will build an uber jar. This type of jar compiles all classes from this project with its dependencies.

To build uber jar

mvn clean package

The command will output .jar file in the target directory.

Run program

cd target
java -cp my-first-dl4j-project-1.0-SNAPSHOT-bin.jar ai.certifai.MyFirstDL4JProject

MyFirstDL4JProject is the class to run which is located in ai.certifai package

Results of AI Model

The AI Model from transfer learning is able to achieve up to 74% accuracy by training with only 250 HD images from each classes.

Training Result

Validation Result

Future Improvements

1. Platform to discuss skin care products
2. Cross geographical skin samples
3. Develop smartphone app
4. More detailed classifier
5. Higher Accuracy

Additional Information

Additional information about this project can be read here.