Computer Vision Projects 

Data: 

• HAM10000 dataset (Human Against Machine with 10000 training images)

• 7 different skin lesion classes

• Source: https://www.kaggle.com/datasets/kmader/skin-cancer-mnist-ham10000/data 

Tools: TensorFlow, Lime

Objectives: 

• Classify the images into 7 different categories using CNN architecture

• Using transfer learning to improve performance

• Using data augmentation for better generalization

• Applying Explainable AI (XAI) tools such as Lime and SHAP for transparency

• Using Generative Adversarial Networks (GAN) to produce synthetic images

Results: The CNN architecture achieves -

• Validation accuracy: 74.69%

• Testing accuracy: 73.19%

Goal: The goal of the project is to design and implement a comprehensive workflow for developing a Convolutional Neural Network (CNN) model and test it on different practical applications.

Tools: TensorFlow, Matplotlib

Workflow: 

• Basic CNN architecture design

•  Dropout Regularization, Batch Normalization

• Callbacks

• Data Augmentation

• Using a Pre-trained CNN model for prediction

• Transfer Learning (Feature extractor, Fine-tuning)

• Interpret model predictions

Applications:

• Hand-drawn Electric circuit Schematic Components prediction

• Malaria parasite detection from blood smears 

Malaria parasite detection:

A CNN architecture is developed to analyze cell images to effectively identify the presence of malaria parasites. The data contains a total of 27,558 images collected from the NIH (National Institutes of Health) website.  Around 95% test accuracy was achieved.