plant disease detection

plant disease detection application built using PyTorch, multi-head attention, MobileNet, ShuffleNet, and a Tkinter GUI for image input. Here's a breakdown of what each part does:

Model Components

MobileNet and ShuffleNet

Pretrained architectures are partially customized by replacing the last layers with convolution or identity layers.

Used as feature extractors within the custom network.

SimpleResidualBlock

A basic residual block using two convolutional layers with skip connection.

ConvBlock

A reusable block of Conv → BatchNorm → ReLU, optionally followed by MaxPool.

ImageClassificationBase

Base class that includes training and validation steps, accuracy calculation, and logging.

? Main Model – ResNet9

This model combines:

A standard CNN backbone (ConvBlock + residual connections).

Feature fusion from:

A customized ShuffleNet (sh)

A customized MobileNet (mobile)

The CNN stream

Three multi-head attention layers, one for each stream, helping the model to focus on relevant spatial features.

The outputs are flattened and concatenated before going through a Linear layer for final classification (38 plant disease classes).

? Model Loading and Inference

Loads a trained model checkpoint: multi (1).pth.

Uses torch.load to restore the model weights.

Applies softmax and gets the class prediction based on the maximum value.

️ GUI with Tkinter

A simple UI to browse an image from the file system.

When an image is selected:

It is resized and transformed into a tensor.

It is passed to the model for prediction.

The predicted label is displayed on the GUI.

️ Classes

labels contains the names of 38 possible plant diseases or healthy categories.

Highlights

Multi-model feature fusion using MobileNet and ShuffleNet.

Multi-head attention for learning richer feature representations.

GUI interface allows practical usage by non-technical users.

Entirely custom model: no pretrained classifier head used.