dglownia/MobileNetV3Large-Bird-Classification-Kaggle

500 Species Bird Classification

by Daniel Glownia

Data

• Size: 224 x 224 x 3 ​
• 500 different bird species with at least 130 train images per species​
• 80% male birds (more colorful) and only 20% female (sex is not labeled)​
• One bird per image​
• Bird takes up 50%+ of pixels​
• Some images include noise like watermarks

Dataset	Image count​
Train	85,085
Test	2,500
Validation	2,500

CNN Implementation

• MobileNetV3 as base model(transfer learning)
• Trained on 100 epochs
• Optimizer: Adam
• Loss: Categorical Cross Entropy

epochs = 100
batch_size = 256

inputs = pretrained_model.input
x = processing_layers(inputs)

x = Dense(256, activation='relu')(pretrained_model.output)
x = Dropout(0.2)(x)
x = Dense(128, activation='relu')(x)
x = Dropout(0.2)(x)
x = Dense(64, activation='relu')(x)
x = Dropout(0.2)(x)


outputs = Dense(500, activation='softmax')(x)

model = Model(inputs=inputs, outputs=outputs)

Results

The following confusion matrix represents the lowest performing classes. Classes with perfect scores were removed.

Dataset	Accuracy
Train	84.87%
Test	92.20%
Validation	93.36%