📘 Wheat Kernel FDK Detection and Analysis

🔍 Overview

This project provides a complete pipeline to automatically detect, classify, and analyze wheat kernels from image plates using a pre-trained YOLOv8 model. It is tailored to identify and quantify kernel traits such as:

• Bleachness (based on average RGB intensity)
• Roughness (based on standard deviation of RGB intensity)
• Crease presence (based on profile shape analysis)

The script performs per-kernel analysis, per-plate summaries, and generates visualizations for statistical interpretation across multiple images.

---

🗂️ Directory Structure

.
├── input_images/              # Folder with test JPG images of plates
├── output/                   # Auto-generated results
│   ├── images/               # Labeled images + kernel grid images
│   ├── individual_kernels_csv/ # CSV files for each plate (per-kernel data)
│   └── plate_level_summary.csv # Final summary: one row per plate
├── wheat_kernel_FDK_Detector.pt  # Trained YOLOv8 model
├── FDK Detection - Poster - Final.pdf  # Poster
└── README.md            # Full pipeline script (this code)

---

⚙️ Setup Instructions

1. Environment Requirements:

• Create a conda environment:

conda create -n cvml python=3.12

• Install pytorch: https://pytorch.org/get-started/locally/ - I installed in Macbook pro M3.

pip3 install torch torchvision torchaudio

• Install jupyterlab:

pip install jupyterlab

• Install opencv:

pip install opencv-contrib-python

• Install ultralytics:

pip install ultralytics

• Install roboflow:

pip install roboflow==1.1.48

2. Input Preparation:
   • Place all .jpg or .jpeg plate images in the input_images/ folder.
   • Make sure the model file wheat_kernel_FDK_Detector.pt is placed in the project root.

---

🚀 How It Works

1. Object Detection with YOLOv8:

   • Kernels are detected and segmented from each image using a confidence threshold of 0.90.

2. Image Masking and Cropping:

   • Each detected kernel is isolated using polygon masks and cropped for analysis.

3. Kernel Rotation:

   • Kernels are aligned horizontally using ellipse fitting to normalize their orientation.

4. Crease Classification:

   • Each rotated kernel is classified as With Crease or Without Crease based on profile length ratio using a smoothed horizontal intensity scan.

5. RGB-Based Feature Extraction:

   • For each kernel:
     • Average Intensity (proxy for bleachness)
     • Standard Deviation (proxy for roughness)
   • Values are rounded and stored for CSV output.

6. CSV Outputs:

   • A .csv file for each plate (individual_kernels_csv/) containing per-kernel stats.
   • A combined summary plate_level_summary.csv with average metrics for each plate.

7. Visual Outputs:

   • Annotated plate images with numbered kernels
   • Grid images of crease vs non-crease kernels

8. Statistical Histograms:

   • Plots generated:
     • % of bleached kernels
     • % of rough kernels
     • % of non-crease kernels
     • Total kernel count per plate

---

📊 Output Description

Per-Kernel CSV (one per plate):

Plate_Name	Kernel_ID	Average_Intensity	Std_Intensity
example1	1	135	48

Summary CSV:

Plate_Name	Average_Intensity	Std_Intensity	Total_Kernels
example1	132.5	47.3	104

---

🧪 Threshold Parameters

• Average Intensity Threshold: 180 – Determines % of bleached kernels
• Standard Deviation Threshold: 65 – Determines % of rough kernels
• These thresholds can be adjusted via:

THRESHOLD_AVG_INTENSITY = 180
THRESHOLD_STD_INTENSITY = 65

---

📈 Visualization

The script generates histograms showing:

• Distribution of bleachness across plates
• Distribution of roughness
• Kernel crease frequency
• Kernel counts per plate

Use these plots to understand phenotypic variability and detect outliers.

---

📞 Contact

Prof. Jessica Rutkoski | jrut@illinois.edu

Prof, Sunoj Shajahan | sunoj@illinois.edu

Seung Wook Yoon | alexyoon9941@gmail.com, swyoon3@illinois.edu