Machine Learning Projects

1. Convolutional Neural Network (CNN)

A foundational deep learning model using convolutional layers to classify leaf diseases, effective for feature extraction. Reference: Krizhevsky et al., 2012

• Utilizes multiple convolutional layers to extract spatial features from leaf images.
• Achieves 85.3% accuracy in detecting diseases like anthracnose and powdery mildew.
• Trained on a dataset of 1000+ leaf images with data augmentation for robustness.

2. DenseNet

Uses dense connectivity between layers to improve gradient flow and feature reuse for leaf disease classification. Reference: Huang et al., 2016

• Each layer receives inputs from all previous layers, enhancing feature propagation.
• Achieves 89.7% accuracy with reduced parameters compared to traditional CNNs.
• Performs well on small datasets due to its dense connectivity pattern.

3. NASNet

Employs neural architecture search to design an efficient network for classifying leaf diseases. Reference: Zoph et al., 2016

• Automatically designs network architecture optimized for leaf disease classification.
• Achieves 87.5% accuracy with optimal computational efficiency.
• Utilizes a search space with recurrent neural networks to find optimal configurations.

4. ResNeXt

Introduces cardinality to ResNet, enhancing feature learning for leaf disease classification. Reference: Xie et al., 2016

• Improves ResNet by adding a cardinality dimension for better feature extraction.
• Achieves 90.2% accuracy in identifying complex disease patterns.
• Reduces overfitting through skip connections and batch normalization.

5. TRESNet

A transformer-based architecture optimized for efficient leaf disease classification. Reference: Ridnik et al., 2020

• Leverages transformer blocks for global feature learning in leaf images.
• Achieves 91.0% accuracy with optimized speed for real-time applications.
• Reduces computational overhead while maintaining high classification accuracy.

6. VGGNet

A deep convolutional network with small filters, effective for classifying leaf diseases. Reference: Simonyan & Zisserman, 2014

• Employs small 3x3 filters in deep layers to capture fine-grained patterns.
• Achieves 86.8% accuracy with high depth (16-19 layers).
• Effective but computationally intensive due to its uniform architecture.

7. Xception

Uses depthwise separable convolutions for efficient leaf disease classification. Reference: Chollet, 2016

• Replaces standard convolutions with depthwise separable convolutions.
• Achieves 88.4% accuracy with improved efficiency.
• Reduces the number of parameters while maintaining high performance.

8. Dense Residual Paths with Attention (DRPA-Net) ⭐

Combines DenseNet's reuse, ResNeXt's cardinality, and attention mechanisms, using transfer learning from pre-trained DenseNet121. Best performing model - 93.5% accuracy. A collaborative design by the team.

• Integrates dense connections and residual paths for enhanced feature learning.
• Uses attention mechanisms to focus on disease-relevant regions.
• Leverages transfer learning from DenseNet121 for improved performance.
• Achieves state-of-the-art 93.5% accuracy in black gram leaf disease classification.

Comparative Results

The 8 models were evaluated on a test set of 150 black gram leaf images:

• DRPA-Net: 93.5% - Best overall performance with attention mechanisms
• TRESNet: 91.0% - Efficient transformer-based approach
• ResNeXt: 90.2% - Strong feature learning with cardinality
• DenseNet: 89.7% - Benefiting from dense connectivity
• Xception: 88.4% - Efficient with depthwise convolutions
• NASNet: 87.5% - Optimized via neural architecture search
• VGGNet: 86.8% - Strong but computationally heavy
• CNN: 85.3% - Solid baseline performance

1. Variational Autoencoder (VAE)

A generative model that learns latent representations of retinal images for enhancement, improving contrast and clarity. Reference: Kingma & Welling, 2016

2. Total Variational De-noising (TVD)

Reduces noise in retinal images by minimizing total variation, preserving edges. Reference: Rudin et al., 1992.

3. Spatial Pyramid Pooling Networks (SPP)

Uses multi-scale feature extraction to enhance retinal images, improving robustness to varying image sizes. Reference: He et al., 2014.

4. Auto Regressive Model

Models pixel dependencies in retinal images for enhancement, focusing on statistical patterns. Reference: Jain, 1981.

5. Normalizing Flows

A generative approach using invertible transformations to enhance retinal images, preserving details. Reference: Rezende & Mohamed, 2015.

6. Energy Based Model (EBM)

Uses energy functions to model and enhance retinal images, focusing on noise reduction. Reference: Du & Mordatch, 2019.

7. Diffusion Model

Applies iterative denoising to enhance retinal images, improving quality for medical analysis. Reference: Ho et al., 2020.

8. Adaptive Local Means CNN (ALM-CNN)

Combines Adaptive Local Means Denoising with a CNN to enhance retinal images, balancing noise reduction and detail preservation. A collaborative design by the team.