FeijõesAI

Automated classification of good and defective beans Python + TensorFlow/Keras pipeline for segmentation, feature extraction and dense neural network classification.

---

📋 Participants

• João Pedro Rocha Senna
• Thiago Tanaka Peczek

---

🤖 AI Name

FeijõesAI - FIA – Dense neural network for binary classification (bad vs. good).

---

🚀 Usage Guide

1. Clone the repository

   git clone https://github.com/Arg0n4ut4/IAFIA.git
   cd IAFIA

2. Create and activate a virtual environment

   python -m venv venv
   source venv/bin/activate    # Linux/macOS
   venv\Scripts\activate       # Windows

3. Install dependencies

   pip install -r requirements.txt

4. Preprocessing Segments and crops each bean from data/raw/good and data/raw/bad → data/processed/

   python src/preprocessing/segment.py data/raw data/processed --size 128

5. Feature extraction

   • Shape:

     python src/features/shape.py data/processed data/features/shape.csv

   • Color:

     python src/features/color.py data/processed data/features/color.csv

   • Texture:

     python src/features/texture.py data/processed data/features/texture.csv

6. Dataset construction Merges shape, color and texture features → splits into 70% training / 30% testing

   python src/datasets/build_dataset.py

   Generates:

   • data/features/train.csv
   • data/features/test.csv

7. Neural Network Training Trains using train.csv (70% of good beans + 70% of bad beans) and internally validates using 20% validation split.

   python src/train.py

   Model saved at data/models/rna.h5.

8. Evaluation (Optional) Run evaluate.py for final metrics using test.csv.

   python src/evaluate.py data/features/test.csv data/models/rna.h5

---

📊 CSV Structure

train.csv / test.csv

Column	Description
area	Bean contour area
perimeter	Contour perimeter
aspect	Bounding box width/height ratio
circularity	4π·area / perimeter²
extent	Area ⁄ (bounding box width·height)
solidity	Area ⁄ convex hull area
hu_1 … hu_7	7 invariant Hu moments (log-transformed)
prop_black	Proportion of near-black pixels (RGB ≤ (30,30,30))
mean_b,mean_g,mean_r	Mean values for B, G and R channels
std_b,std_g,std_r	Standard deviation of B, G and R channels
prop_dark_otsu	Proportion of dark pixels in V channel (HSV + Otsu)
glcm_contrast	GLCM contrast
glcm_homogeneity	GLCM homogeneity
glcm_energy	GLCM energy
glcm_correlation	GLCM correlation
lbp_0 … lbp_9	LBP histogram (10 bins, “uniform” method)
hog_0 … hog_N	HOG feature vector (gradient orientations)
label	Class: 0 = bad, 1 = normal

Note:

• train_reduced.csv and test_reduced.csv contain only the 10 most important features (calculated using RandomForestClassifier.feature_importances_) + label.

---

🧪 Methodology

1. Preprocessing

   • Grayscale conversion + Otsu thresholding
   • Morphological operations (open)
   • Main contour detection + crop + resize (128×128)

2. Feature Extraction

   • Shape: area, perimeter, aspect ratio, circularity, extent, solidity, Hu moments
   • Color: dark pixel proportion, RGB statistics, dark V-channel proportion (HSV + Otsu)
   • Texture: GLCM (contrast, homogeneity, energy, correlation), LBP (10 bins), HOG

3. Feature Selection

   • Training a RandomForestClassifier(n_estimators=100) using train.csv
   • Feature importance ranking and retention of the top 10 features
   • Generation of train_reduced.csv and test_reduced.csv

4. Dense Neural Network

   • Fully-connected layers: 128 → 64 → 32
   • BatchNormalization + Dropout(0.3)
   • Sigmoid output (binary_crossentropy)
   • Metrics: accuracy, AUC
   • Callbacks: EarlyStopping (patience=10), ModelCheckpoint

---

📈 Results and Metrics

Metric	Full Model	Reduced Model
Accuracy	90.20 %	89.71 %
F1-score (average)	0.9091	0.9041
ROC-AUC	0.9628	0.9511
Precision (bad)	0.88	0.87
Recall (bad)	0.91	0.91
Precision (normal)	0.93	0.93
Recall (normal)	0.89	0.88

---

---

FeijõesAI

Classificação automatizada de feijões bons e defeituosos Pipeline em Python + TensorFlow/Keras para segmentação, extração de atributos e rede neural densa.

---

📋 Participantes

• João Pedro Rocha Senna
• Thiago Tanaka Peczek

---

🤖 Nome da IA

FeijõesAI - FIA – Rede neural densa para classificação binária (ruim vs. bom).

---

🚀 Passo a passo de uso

1. Clone o repositório

   git clone https://github.com/Arg0n4ut4/IAFIA.git
   cd IAFIA

2. Crie e ative um ambiente virtual

   python -m venv venv
   source venv/bin/activate    # Linux/macOS
   venv\Scripts\activate       # Windows

3. Instale as dependências

   pip install -r requirements.txt

4. Pré-processamento Segmenta e recorta cada grão de data/raw/bons e data/raw/ruins → data/processed/

   python src/preprocessing/segment.py data/raw data/processed --size 128

5. Extração de features

   • Forma:

     python src/features/shape.py data/processed data/features/shape.csv

   • Cor:

     python src/features/color.py data/processed data/features/color.csv

   • Textura:

     python src/features/texture.py data/processed data/features/texture.csv

6. Construção do dataset Junta shape, color e texture → divide 70% treino / 30% teste

   python src/datasets/build_dataset.py

   Gera:

   • data/features/train.csv
   • data/features/test.csv

7. Treinamento da RNA Treina com train.csv (70% dos bons + 70% dos ruins) e valida internamente (20% de validação).

   python src/train.py

   Modelo salvo em data/models/rna.h5.

8. Avaliação (Opcional) Rode evaluate.py para métricas finais no test.csv.

   python src/evaluate.py data/features/test.csv data/models/rna.h5

---

📊 Estrutura dos CSVs

train.csv / test.csv

Coluna	Descrição
area	Área do contorno do grão
perimeter	Perímetro do contorno
aspect	Razão largura/altura da bounding box
circularity	4π·area / perímetro²
extent	Área ⁄ (largura·altura) da bounding box
solidity	Área ⁄ (área do convex hull)
hu_1 … hu_7	7 momentos de Hu invariantes (log-transformados)
prop_black	Proporção de pixels quase-pretos (RGB ≤ (30,30,30))
mean_b,mean_g,mean_r	Média dos canais B, G, R
std_b,std_g,std_r	Desvio-padrão dos canais B, G, R
prop_dark_otsu	Proporção de pixels escuros no canal V (HSV + Otsu)
glcm_contrast	Contraste da matriz GLCM
glcm_homogeneity	Homogeneidade da matriz GLCM
glcm_energy	Energia da matriz GLCM
glcm_correlation	Correlação da matriz GLCM
lbp_0 … lbp_9	Histograma LBP (10 bins, método “uniform”)
hog_0 … hog_N	Vetor HOG (orientações de gradiente)
label	Classe: 0 = ruim, 1 = normal

Observação:

• train_reduced.csv e test_reduced.csv contêm apenas as 10 features mais importantes (calculadas via RandomForestClassifier.feature_importances_) + label.

---

🧪 Metodologias

1. Pré-processamento

   • Conversão para escala de cinza + Otsu
   • Operações morfológicas (open)
   • Detecção de contorno principal + crop + resize (128×128)

2. Extração de atributos

   • Forma: área, perímetro, razão, circularidade, extent, solidity, momentos de Hu
   • Cor: proporção de pixels escuros, estatísticas RGB, proporção escura em V (HSV+Otsu)
   • Textura: GLCM (contrast, homog., energy, corr.), LBP (10 bins), HOG

3. Seleção de features

   • Treinamento de RandomForestClassifier(n_estimators=100) em train.csv
   • Ordenação por importância e retenção das 10 top features
   • Geração de train_reduced.csv e test_reduced.csv

4. Rede Neural Densa

   • Camadas fully-connected: 128 → 64 → 32
   • BatchNormalization + Dropout(0.3)
   • Saída Sigmoid (binary_crossentropy)
   • Métricas: accuracy, AUC
   • Callback: EarlyStopping (patience=10), ModelCheckpoint

---

📈 Resultados e Métricas

Métrica	Modelo Completo	Modelo Reduzido
Acurácia	90,20 %	89,71 %
F1-score (médio)	0,9091	0,9041
ROC-AUC	0,9628	0,9511
Precision (ruim)	0,88	0,87
Recall (ruim)	0,91	0,91
Precision (normal)	0,93	0,93
Recall (normal)	0,89	0,88