Early detection of Salmonella spp. in poultry is crucial for improving the productivity of poultry companies in the country, as it not only reduces the mortality in their farms and maintains the profitability and reputation of the companies, but also guarantees the production of safe food for the population. However, the response capacity and cost of traditional methods limits the microbiological performance of these companies.
In this project, different architectures of Convolutional Neural Networks (CNN) were evaluated for the detection of salmonella spp. in images of poultry feces. A set of 5029 images taken in Africa between 2020 and 2021 was used, and they were classified as "Healthy" or "Salmonella." After data cleaning and preprocessing, it was found that the implementation of the L2 regularizer and a Dropout layer (0.5) in combination with the Adam optimizer (lr = 0.0001) are the optimal starting parameters for the transfer learning technique in the pre-trained VGG16 and ResNet50v2 models.
The results showed that both models outperformed the baseline model in terms of accuracy and recall, with VGG16 achieving a precision of 98.42% and ResNet50v2 of 93.67%. In conclusion, the effectiveness of computer vision and deep learning techniques for the detection of Salmonella spp. was demonstrated, which provides a solid foundation for future research with images taken in Colombia.
| File | Summary |
|---|---|
01_data_prep_EDA.ipynb |
In this notebook, 14 convolutional neural network (CNN) models are designed and compared to classify images into two categories: Healthy or Salmonella. The models are compared in terms of accuracy, using different combinations of optimizers (Adam and RMSprop), regularizers (L1, L2, L1_L2, and none), and Dropout layers. The performance of each model is evaluated using confusion matrices and performance metrics such as accuracy and recall. |
02_baseline.ipynb |
In this notebook, 14 convolutional neural network (CNN) models are designed and compared to classify images into two categories: Healthy or Salmonella. The models are compared in terms of accuracy, using different combinations of optimizers (Adam and RMSprop), regularizers (L1, L2, L1_L2, and none), and Dropout layers. The performance of each model is evaluated using confusion matrices and performance metrics such as accuracy, recall and auc. |
03_transfer_learning_VGG16.ipynb 04_transfer_learning_ResNet50v2.ipynb |
This Notebooks cover: 1. Data Preprocessing: Preparing images with techniques like data augmentation. 2. Initial Training: Extract images characteristics using the VGG16 or ResNet50v2 network on the training set and evaluating on the validation set. 3. Fine-tuning: Unfreezing specific layers for improved Salmonella classification. 4. Model Evaluation: Assessing the model's performance with confusion matrix visualization and metric comparison, including the test set. |
- Usar ramas diferentes para cada issue
- Usar la guía de estilos de Python de Google
- Mantener los commits cortos y con descripción