Project 1: SA-CNN – Partially Adaptive Optimization driven Spatially Attentive Convolutional Neural Network for Hand-written Signature Verification
Project Overview:
- Objective: Develop an efficient and accurate signature verification system to combat signature forgery in secure transactions like banking, event sign-ups, and official documents.
- Industry/Domain: Smart Security Systems, Biometric Authentication
Key Technologies Used:
- Programming Languages: Python
- Libraries/Frameworks: TensorFlow, Keras
- Techniques: Convolutional Neural Networks (CNN), Spatial Attention Mechanism, Custom Adam Optimizer
Challenges Faced:
- Signature verification models were often too complex, impacting speed and efficiency.
- Developed a model that balances accuracy with fast processing speed.
Project Details:
- Problem Statement: Accurate verification of handwritten signatures to prevent unauthorized transactions.
- Solution: Proposed the SA-CNN model that incorporates spatial attention and adaptive optimization to improve signature recognition and classification speed.
- Results: Achieved a remarkable 99.95% accuracy with reduced iterations, outperforming standard models on benchmark datasets (CEDAR).
- Metrics: Accuracy: 99.95%, Equal Error Rate (EER), Mean-Squared Error (MSE), Area Under Curve (AUC-ROC)
Skills Demonstrated:
- Technical Skills: AI, Machine Learning, Signature Verification, CNN, Data Processing
- Soft Skills: Problem-solving, Research, Innovation
Project 2: CF-CNN – Channel Focused Convolutional Neural Network for Interpretable Ophthalmic Diagnosis
Project Overview:
- Objective: Automate the diagnosis of eye diseases using Optical Coherence Tomography (OCT) images with high accuracy and interpretability.
- Industry/Domain: Medical Imaging, Ophthalmology
Key Technologies Used:
- Programming Languages: Python
- Libraries/Frameworks: TensorFlow, Keras
- Techniques: Convolutional Neural Networks (CNN), Dynamic Convolution, Spatial Attention, Explainable AI (XAI)
Challenges Faced:
- Existing models were either too complex or lacked interpretability.
- Needed to balance accuracy with the ability to explain the AI’s decision-making process.
Project Details:
- Problem Statement: Accurate classification of OCT images into categories like Normal, Choroidal Neovascularization (CNV), Diabetic Macular Edema (DME), and Drusen.
- Solution: Developed the CF-CNN model that uses dynamic convolution and spatial attention, with an explainability layer that highlights key features for diagnosis.
- Results: Achieved an accuracy of 96.68% with fast classification and clear decision explanations using XAI techniques.
- Metrics: Accuracy: 96.68%, Dice Coefficient, Mean Absolute Error (MAE)
Skills Demonstrated:
- Technical Skills: Medical Image Analysis, AI, Machine Learning, CNN, Explainable AI (XAI)
- Soft Skills: Innovation, Analytical Thinking, Communication
Project 3: DA-CNN – Explainable AI driven Dual Attentive Convolutional Neural Network for Retinal Disease Classification
Project Overview:
- Objective: Provide an AI solution for classifying retinal diseases in OCT images with interpretability, using a dual-attention CNN model.
- Industry/Domain: Ophthalmology, Medical Imaging
Key Technologies Used:
- Programming Languages: Python
- Libraries/Frameworks: TensorFlow, Keras
- Techniques: CNN, Dual Attention Mechanism, Explainable AI (XAI)
Challenges Faced:
- Balancing accuracy, speed, and interpretability in a single system for medical diagnostics.
- Ensuring that the model’s decisions were explainable to medical professionals.
Project Details:
- Problem Statement: Automate retinal disease classification into categories such as Normal, Diabetic Macular Edema (DME), Drusen, and CNV.
- Solution: Introduced DA-CNN that uses dual-attention and separable dynamic convolution for feature extraction, while also providing explanations of its decisions using XAI.
- Results: The model achieved an optimal accuracy of 97.4%, surpassing other models on key evaluation metrics.
- Metrics: Accuracy: 97.4%, MAE, Dice Coefficient, AUC-ROC, Sensitivity, Specificity
Skills Demonstrated:
- Technical Skills: Machine Learning, Medical Imaging, Retinal Disease Classification, XAI
- Soft Skills: Problem-solving, Research, Presentation Skills
Project 4: Satellite Image Classification using Explainable AI-based CNN for Remote Sensing
Project Overview:
- Objective: Automate the classification of satellite images for remote sensing applications like urban planning and environmental monitoring.
- Industry/Domain: Remote Sensing, Environmental Science, Urban Planning
Key Technologies Used:
- Programming Languages: Python
- Libraries/Frameworks: TensorFlow, Keras
- Techniques: CNN, Dynamic Convolution, Dual Attention Mechanism, Explainable AI (XAI)
Challenges Faced:
- Existing models lacked simplicity, scalability, and explainability for satellite image classification.
- Needed a system that was fast, accurate, and interpretable.
Project Details:
- Problem Statement: Classify satellite images into categories like green area, desert, clouds, and water for environmental analysis.
- Solution: Developed a CNN model using separable dynamic convolution and channel-split dual attention for enhanced feature extraction, and utilized XAI techniques to explain classification decisions.
- Results: The model achieved a high accuracy of 97.69%, outperforming existing models in classification speed and interpretability.
- Metrics: Accuracy: 97.69%, Precision, Recall, F1-Score
Skills Demonstrated:
- Technical Skills: Satellite Image Classification, Remote Sensing, AI, CNN, XAI
- Soft Skills: Technical Research, Scalability, Communication