Anomaly Detection on Grape Images Using CNNs

The European Canopies initiative: advancing human-robot collaboration in precision agriculture

🍇 Project Overview

This project is part of the European Canopies initiative, aimed at enhancing human-robot collaboration in precision agriculture. The goal is to automatically detect anomalies in grape berries using Convolutional Neural Networks (CNNs), particularly a LeNet-based architecture, and compare its performance with a benchmark study.

The system focuses on automatically distinguishing between healthy and damaged grapes to improve crop quality monitoring and plant health assessment, contributing to more efficient and sustainable agricultural practices.

📘 Problem & Solution

Problem: Automatically distinguish between healthy and damaged grapes to improve crop quality and plant health monitoring
Solution: A CNN trained on a custom dataset of grape images, optimized via hyperparameter tuning and evaluated against a reference paper
Tools: Python, PyTorch, Google Colab, Matplotlib, Seaborn

🗂️ Dataset

The dataset is composed of RGB patches at multiple zoom levels (focus on 1.5x), each containing a single grape.

Organization:

good/ (healthy grapes)
bad/ (damaged grapes)

To handle dataset imbalance (1:4 bad to good), image augmentation was applied via 90°, 180°, and 270° rotations, effectively balancing the dataset and improving model generalization.

🧠 State of the Art

The benchmark paper used a LeNet CNN and compared it with a deeper ResNet50 (pretrained and from scratch). Despite the complexity of ResNet, LeNet showed greater stability and accuracy in the specific binary classification task.

Performance from the benchmark paper:

Training accuracy: 96.4%
Validation accuracy: 97%
Precision: 96.26%
Recall: 93.34%

🛠️ Implementation & Optimization

Technical Configuration:

Weight Initialization: Xavier (Glorot)
Loss Function: BCEWithLogitsLoss
Optimizers Compared: SGD vs Adam → Chosen: SGD

Hyperparameter Tuning (Grid Search):

Learning rate: 0.001
Momentum: 0
Weight decay: 0.001

Model saving via torch.save() was used to handle Google Colab time constraints and ensure reproducibility.

🔍 Results

Performance Metrics:

Validation Accuracy: ~80%
Precision: 75%
Recall: 85%
Training stopped at epoch 70 to prevent overfitting

✅ Strengths:

Effective in detecting damaged berries with darker hues and rough textures
Good generalization with proper data augmentation

⚠️ Weaknesses:

False positives: Blurry or shiny healthy grapes misclassified as damaged
False negatives: Subtly damaged grapes within healthy clusters

🧾 Conclusions

While the model performed reasonably well, it did not match the reference paper due to dataset quality issues:

Mixed-class patches (both healthy and damaged berries in single images)
Lack of precise centering, unlike the benchmark dataset
Inconsistent lighting and background conditions

📌 Main Insight:

Better image acquisition and labeling are crucial for achieving state-of-the-art results in agricultural computer vision applications. The quality of training data often matters more than model complexity.

🔮 Future Improvements:

Improved data collection with consistent lighting and positioning
More sophisticated data augmentation techniques
Integration with real-time robotic systems for field deployment

BibTeX

@misc{decarlo2023grapes,
  title={Anomaly Detection on Grape Images Using CNNs},
  author={De Carlo, Andrea},
  year={2023},
  institution={Sapienza University of Rome},
  type={Bachelor's Thesis},
  note={European Canopies Initiative}
}