Leveraging advanced image processing and deep learning, this project focuses on CNNs and the Keras API for image processing and regression tasks related to plant images, particularly weed species from Plant Seedlings dataset"I worked on a subset". The project involves data preparation, basic transfer learning using the VGG-16 model, classification, and regression networks. Regularization methods are applied to improve the model, and discussions on overfitting and the impact of regularization are included. The submission requires a Jupyter file containing the solution, and late submissions are not allowed. It's contributes to understanding CNNs, transfer learning, and handling small training data. This project holds significance within my Master's in Computer Vision at uOttawa (2023).
- Required libraries: scikit-learn, pandas, matplotlib.
- Execute cells in a Jupyter Notebook environment.
- The uploaded code has been executed and tested successfully within the Google Colab environment.
The dataset comprises 4 classes with 250 images each, divided into training,and testing sets, images size are differnet: Cleavers, Common Chickweed, Maize, Shepherd’s Purse,
-
Data Preparation:
-
Uploaded a subset of the dataset from Google Drive.
-
Extracted the dataset and organized it into 70% training, 15% validation, and 15% testing sets.
- Traning Set
- Validation Set
- Testing Set
- Traning Set
-
Loaded the data, resized images to 32x32 pixels, and created DataFrames for each set.
Training Data Size: 700 Training Data Label Counts: Shepherds_Purse 175 Common_Chickweed 175 Cleavers 175 Maize 175 Name: Label, dtype: int64 Size of the Images in Training Data: (32, 32, 3) ---------------------------------------------------------------- Validation Data Size: 148 Validation Data Label Counts: Shepherds_Purse 37 Common_Chickweed 37 Cleavers 37 Maize 37 Name: Label, dtype: int64 Size of the Images in Validation Data: (32, 32, 3) ---------------------------------------------------------------- Test Data Size: 152 Test Data Label Counts: Shepherds_Purse 38 Common_Chickweed 38 Cleavers 38 Maize 38 Name: Label, dtype: int64 Size of the Images in Test Data: (32, 32, 3) ----------------------------------------------------------------
-
-
Classification Network (Transfer Learning):
- Used the first 2 blocks of VGG-16 model for transfer learning.
- Modified the model by adding custom layers for classification.
# Add custom layers x = Conv2D(256, (3, 3), activation='relu', padding='same')(x) x = MaxPooling2D((2, 2))(x) x = Conv2D(128, (3, 3), activation='relu', padding='same')(x) x = MaxPooling2D((2, 2))(x) x = Flatten()(x) outputs = Dense(4, activation='softmax')(x) # Output layer for 4 classes # Create the custom model classification_model = Model(inputs=vgg_model.input, outputs=outputs)
- One-hot encoded the labels.
- Trained the classification model, monitored convergence, and visualized learning curves.
batchSize = 64 nEpochs = 100 # Compile the model classification_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy']) # Train the model history = classification_model.fit(X_train, y_train_k, batch_size=batchSize, epochs=nEpochs, verbose=1, validation_data=(X_valid, y_valid_k))
- Plotted and analyzed the confusion matrix for training, validation, and testing datasets.
-
Regression Network (Transfer Learning):
-
Loaded bounding box dimensions from the
.jsonfile. -
Normalized height and width values.
-
Split the data into 70% training, 15% validation, and 15% testing sets.
Training Data Size: 700 Training Data Label Counts: Shepherds_Purse 175 Common_Chickweed 175 Cleavers 175 Maize 175 Name: Label, dtype: int64 ---------------------------------------------------------------- Validation Data Size: 148 Validation Data Label Counts: Shepherds_Purse 37 Common_Chickweed 37 Cleavers 37 Maize 37 Name: Label, dtype: int64 ---------------------------------------------------------------- Test Data Size: 152 Test Data Label Counts: Shepherds_Purse 38 Common_Chickweed 38 Cleavers 38 Maize 38 Name: Label, dtype: int64
-
Used VGG-16 for transfer learning with custom layers for regression.
# Add custom layers x_regression = Conv2D(256, (3, 3), activation='relu', padding='same')(x_regression) x_regression = MaxPooling2D((2, 2))(x_regression) x_regression = Conv2D(128, (3, 3), activation='relu', padding='same')(x_regression) x_regression = MaxPooling2D((2, 2))(x_regression) x_regression = Flatten()(x_regression) height_output = Dense(1, activation='linear', name='height')(x_regression) width_output = Dense(1, activation='linear', name='width')(x_regression) # Create the custom regression model regression_model = Model(inputs=regression_vgg_model.input, outputs=[height_output, width_output])
-
Trained the regression model, monitored convergence, and visualized learning curves.
batchSize = 64 nEpochs = 100 # Compile the model regression_model.compile(optimizer='adam', loss='mean_squared_error', metrics=['mean_squared_error']) # Train the regression model results =regression_model.fit(X_train_regression, [y_train_height, y_train_width], epochs=nEpochs, validation_data= (X_valid_regression, [y_valid_height, y_valid_width]))
-
Calculated mean squared error and mean absolute error for training, validation, and testing datasets.
22/22 [==============================] - 0s 4ms/step Mean Squared Error for height - Train: 0.002856253375326049, width - Train: 0.003164909554132075 Mean Absolute Error for height - Train: 0.04336496062917911, width - Train: 0.04329592842347164 5/5 [==============================] - 0s 4ms/step Mean Squared Error for height - Validation: 0.09055006138348325, width - Validation: 0.06981748160195345 Mean Absolute Error for height - Validation: 0.2218270389548888, width - Validation: 0.20934197684151834 5/5 [==============================] - 0s 4ms/step Mean Squared Error for height - Test: 0.07094346629570776, width - Test: 0.08139776182780212 Mean Absolute Error for height - Test: 0.2150076942617718, width - Test: 0.22207330307667558
-
-
Model Improvement (Classification Network):
- Modified the VGG-16 model by adding extra Keras layers and Introduced regularization techniques such as Batch
Normalization and Dropout.# Add custom layers with regularization x_new = Conv2D(256, (3, 3), activation='relu', padding='same')(x_new ) x_new = BatchNormalization()(x_new ) # Batch Normalization layer x_new = MaxPooling2D((2, 2))(x_new) x_new = Conv2D(128, (3, 3), activation='relu', padding='same')(x_new) x_new = BatchNormalization()(x_new) # Batch Normalization layer x_new = MaxPooling2D((2, 2))(x_new) x_new = Flatten()(x_new) x_new = Dropout(0.5)(x_new) # Dropout layer with a dropout rate of 0.5 outputs = Dense(4, activation='softmax')(x_new) # Output layer for 4 classes # Create the model new_custom_model = Model(inputs=new_vgg_model.input, outputs=outputs)
- Trained the improved classification model, monitored convergence, and visualized learning curves.
batchSize = 35 nEpochs = 100 # Compile the model new_custom_model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy']) # Train the model newModel = new_custom_model.fit(X_train, y_train_k, batch_size=batchSize, epochs=nEpochs, verbose=1, validation_data=(X_valid, y_valid_k))
- Plotted and analyzed the confusion matrix for training, validation, and testing datasets.
!
- Modified the VGG-16 model by adding extra Keras layers and Introduced regularization techniques such as Batch