Summary: Image classification by disease recognition on leaves.

There are 4 distinct parts in this project, 01. Distribution, 02. Augmentation, 03. Transformation, and 04. Classification.

Download image dataset and generate distribution chart image

usage: 01.Distribution.py [-h] directories [directories ...]

A program to analyze plant images and generate charts.

positional arguments:
  directories  The directories to store extracted images and save the charts (ex: 01.Distribution apple)

options:
  -h, --help   show this help message and exit

python3  01.Distribution.py  apple  grape

Augment unbalanced image dataset

usage: 02.Augmentation.py [-h] [file_path]

A program to augment images samples by applying 6 types of transformation.

positional arguments:
  file_path   Image file path to transform to 6 different types.

options:
  -h, --help  show this help message and exit

python3  02.Augmentation.py

Save transformed image plots

usage: 03.Transformation.py [-h] -src [SRC_PATH] [-dst [DST_PATH]] [-gaussian] [-mask] [-roi] [-analyze] [-pseudo] [-hist]

A program to display image transformation.

options:
  -h, --help            show this help message and exit
  -src [SRC_PATH], --src_path [SRC_PATH]
                        Image file path.
  -dst [DST_PATH], --dst_path [DST_PATH]
                        Destination directory path.
  -gaussian, --gaussian_blur
                        Gaussian Transform
  -mask                 Mask Transform
  -roi, --roi_objects   Roi Transform
  -analyze, --analyze_object
                        Analyze Transform
  -pseudo, --pseudolandmarks
                        Psudolandmark Transform
  -hist, --color_histogram
                        Color histogram Transform

python3  03.Transformation.py  -src [SRC_PATH] -dst [DST_PATH]

Print the accuracy on validation dataset

usage: 04.Classification.py [-h] [folder_path]

A program to classify a type of leaf from validation set.

positional arguments:
  folder_path  Image folder path.

options:
  -h, --help   show this help message and exit

python3  04.Classification

The model is designed to classify leaf diseases based on images of leaves. The model is implemented using Pytorch and consists of 4 convolutional layers followed by max pooling, along with 2 fully connected layers. The final output is produced using a softmax function for multi-class classification.

1. Input layer

   • Input: Leaf images with a shape of (256, 256, 3) corresponding to 256 x 256 RGB images.

2. Convolutional layers

   • Conv Layer 1
     • Input channels: 3 (RGB)
     • Output channels: 32
     • Kernel size: 3 x 3
     • Activation function: ReLU
     • Max Pooling: 2 x 2
   • Conv Layer 2
     • Input channels: 32
     • Output channels: 64
     • Kernel size: 3 x 3
     • Activation function: ReLU
     • Max Pooling: 2 x 2
   • Conv Layer 3
     • Input channels: 64
     • Output channels: 128
     • Kernel size: 3 x 3
     • Activation function: ReLU
     • Max Pooling: 2 x 2
   • Conv Layer 4
     • Input channels: 128
     • Output channels: 256
     • Kernel size: 3 x 3
     • Activation function: ReLU
     • Max Pooling: 2 x 2

3. Fully connected layers

   • FC Layer 1
     • Input: Flattened tensor from the previous convolutional layers (256 * 14 * 14 = 50176 units)
     • Output: 512 units
     • Activation function: ReLU
     • Dropout: 0.5
   • FC Layer 2
     • Input: 512 units
     • Output: NUM_CLASSES units (representing the number of disease classes)
     • Activation function: Softmax

There are 2 distinct leaf types; apple and grape, each of which consists of 4 labels.

Apple Image Distribution	Grape Image Distribution

The following 6 image augmentation techniques are applied to one single-leaf image labeled apple black rot.

Brightness	Contrast	Flip	Perspective	Rotate	Saturation

The following 6 image transformation techniques are applied to one single-leaf image labeled apple black rot.

Mask	Gaussian Blur	Roi objects	Analyze object	Pseudolandmarks

Color Histogram

To visualize the learning curves using tensorboard, execute the following command.

tensorboard --logdir runs

We have 10 test images and the model has 100% accuracy

• Youtube Coursera CNN