Grape Leaf Disease Detection with KANConv2D

Description

This project used a custom convolutional layer, KANConv2D, to build a convolutional neural network (CNN) for detecting Esca disease in grape leaves. The KANConv2D layer incorporates kernel adaptive networks to enhance the performance of traditional Conv2D layers.

For more on KAN see: https://arxiv.org/pdf/2406.13155

Final accuracy was 84.17%

Features

Custom KANConv2D layer with kernel adaptive networks
Data augmentation for robust training
Early stopping to prevent overfitting
Class weighting to handle imbalanced datasets

Model Architecture

The CNN model consists of:

Multiple KANConv2D layers
Max pooling layers
Dense and dropout layers
Binary output with sigmoid activation

Training

The model is compiled with the Adam optimizer, binary cross-entropy loss, and accuracy as the metric. Class weights are adjusted to reduce false negatives for Esca detection. The model is trained with early stopping to restore the best weights based on validation loss.