Poultry-Meat-Freshness-Classification-with-Transfer-Learning-using-ResNet-Architecture

The utilization of the ResNet architecture and Transfer Learning methods to classify Poultry Meat Freshness images into two categories, namely Fresh and Rotten.

About The Project

• In this repository, we leverage the power of the Residual Networks (ResNet) architecture and Transfer Learning techniques to accurately classify the freshness of poultry meat. By utilizing pre-trained ResNet weights, sourced from ImageNet, our model gains the ability to make fine-grained distinctions in poultry meat freshness.
• The pre-trained weights from the ImageNet dataset, which includes a wide range of object categories, are used to enhance the model's ability to recognize and classify fresh and rotten poultry meat images.
• This repository provides a comprehensive implementation of the classification process and serves as a valuable resource for exploring the world of Transfer Learning with ResNet architecture.
• Key Features
  • Implementation of Transfer Learning with ResNet architecture.
  • Classification of poultry meat freshness into different categories.
  • Utilizes a pre-trained model to enhance classification accuracy.
  • Implement hyperparameter tuning using Grid Search to obtain the best model/ set of parameters.
  • Helps in maintaining food safety and quality standards.

Technology Used

• Python
• Numpy
• Pandas
• Matplotlib
• Scikit-learn
• Keras
• Tensorflow

Dataset Used

• In this project, we utilize sets of images depicting fresh and spoiled poultry meat sourced from the Fresh and rotten poultry meat datasets, accessible at: Fresh and rotten poultry meat datasets | Kaggle.
• The amount of data used in this project:
  • Training data:
    • Fresh (Segar): 500 images
    • Rotten (Busuk): 500 images
  • Testing data
    • Fresh (Segar): 150 images
    • Rotten (Busuk): 150 images
• Sample Images
  • Fresh (Segar)

    1	2	3	4	5

  • Rotten (Busuk)

    1	2	3	4	5

Workflow

• Data Preparation
• Label Encoding
• Data Preprocessing & Data Augmentation
• Data Splitting
• Model Building
• Model Training & Hyperparameters Tuning
• Model Testing & Evaluation

Algorithms/ Methods

• This project applies Transfer Learning methods by utilizing the ResNet architecture with pre-trained weights sourced from ImageNet, encompassing around 1000 object categories.
• The training process uses a fine-tuned method, which allows all of the layers to update their weights during the training process.
• Parameters (Fixed):
  • Batch = 10 (100 steps per epoch, because the number of training data is 1000)
  • Optimizer = Mini Batch Gradient Descent
  • Loss/ Cost Function = Binary Cross Entropy
  • Metrics = Accuracy
• Hyperparameters (Fine Tuned with Grid Search):
  • Epoch:
    • 100
    • 200
  • Learning rate:
    • 0.0001
    • 0.0003
    • 0.0007
  • Momentum:
    • 0.0
    • 0.9

Model Evaluation

Graph of Epoch & Accuracy (100 epoch)	Graph of Epoch & Accuracy (200 epoch)
Graph of Loss & Accuracy (100 epoch)	Graph of Loss & Accuracy (200 epoch)

• Validation accuracy scores at Epoch 100:
  • resnet_lr1e-4_m0_e100: 92.33
  • resnet_lr3e-4_m0_e100: 95.33
  • resnet_lr7e-4_m0_e100: 94.67
  • resnet_lr1e-4_m09_e100: 93.67
  • resnet_lr3e-4_m09_e100: 92.67
  • resnet_lr7e-4_m09_e100: 92.0
• Validation accuracy scores at Epoch 200:
  • resnet_lr1e-4_m0_e200: 93.0
  • resnet_lr3e-4_m0_e200: 95.0
  • resnet_lr7e-4_m0_e200: 95.67
  • resnet_lr1e-4_m09_e200: 91.0
  • resnet_lr3e-4_m09_e200: 91.67
  • resnet_lr7e-4_m09_e200: 90.67

Publication

• Cendani, L. M., Pangestu, M. A., & Muria, F. (2021). "Classification of Freshness Quality of Broiler Chicken Breast with Transfer Learning Using ResNet, Inception, and Xception Architectures". (unpublished)

Contributors

• Linggar Maretva Cendani - [email protected] (ResNet)
• Michael Axel Pangestu (Inception)
• Fatah Muria (Xception)

License

This project is licensed under the MIT License - see the LICENSE file for details

Acknowledgments

• Fresh and rotten poultry meat datasets: https://www.kaggle.com/calvinsama/fresh-and-rotten-poultry-meat-datasets.