Bird Identification App

Project Overview

Computer vision app for birdwatchers to identify 7 Dutch bird species via photos. MobileNet transfer learning model trained on custom dataset of 2800+ images collected with Google Image Scraper. Completed full product development cycle from data collection to Figma prototype with A/B testing. The project combines technical model development with user-centered design for a mobile app.

Challenges

Working with limited training data (400 images per species) required smart use of transfer learning and data augmentation. Visually similar bird species like blue tit and great tit were difficult to distinguish and required careful feature extraction. Manual data curation of all scraped images to guarantee data quality was time-consuming. Finding balance between model accuracy and mobile deployment efficiency (MobileNet vs heavier architectures).

Results

Model achieved 95.2% test accuracy with consistently high scores (92-98% precision, 92-100% recall) for all 7 bird species. Transfer learning with MobileNet made these results possible with relatively little data. Complete Figma prototype developed with camera interface, spotting hotspots map, and species directory. A/B test with 10 users showed preference for colorful design variant. Project demonstrates complete ML pipeline from data collection to production-ready model with user-friendly interface.

95%

Test Accuracy

7

Bird Species

2800+

Training Images

10

User Tests (A/B)

Model Development: Transfer Learning Pipeline

🧠 From Data Collection to Trained Model

The project demonstrates the complete machine learning lifecycle: from collecting and annotating custom data via Google Image Scraper to a production-ready model with 95% accuracy. By combining transfer learning with MobileNet with data augmentation, excellent results were achieved with limited training data.

Classification of 7 Dutch Bird Species

🐦 Blackbird

🐦 Blue Tit

🐦 Dunnock

🐦 Chaffinch

🐦 Great Tit

🐦 House Sparrow

🐦 Robin

Data Collection & Preprocessing

Dataset creation: Google Image Scraper used to collect ~400 images per bird species
Manual selection: All scraped images manually screened for quality and correctness
Data augmentation: Dataset doubled via rotation (30°), shifts (20%), shear, zoom, and horizontal flip
Normalization: MinMaxScaler for pixel value normalization, MobileNet preprocessing for transfer learning
Train/Val/Test split: 70/20/10 distribution for robust model evaluation

Model Architecture: MobileNet Transfer Learning

Why MobileNet?

MobileNet is specifically designed for mobile deployment and offers an ideal balance between accuracy and efficiency. The pre-trained weights from ImageNet provide a strong foundation for feature extraction, achieving better performance even with a relatively small custom dataset of ~400 images per class.

Architecture Details

Base model: MobileNetV2 pre-trained on ImageNet
Input: 224x224x3 RGB images (standard MobileNet input size)
Feature extractor: MobileNet convolutional base (frozen weights) generates 7x7x1024 features
Custom classifier: Dense layers (512→256→128 neurons) with ReLU activation and Dropout (0.5) for regularization
Output: 7-class softmax layer for probabilistic bird species classification

Training Strategy

Optimizer: Adam with learning rate 1e-3 for fast and stable convergence
Batch size: 32 samples for good generalization without overfitting
Early stopping: Patience of 5 epochs prevents overfitting and saves training time
Iterative improvement: Multiple model versions tested (V1-V5) with different architectures and data preprocessing

Model Performance & Results

95.2%

Test Accuracy

0.17

Test Loss

92-98%

Per-Class Precision

92-100%

Per-Class Recall

Detailed Classification Report

The model achieves consistently high scores across all bird species:

Dunnock: 93% precision, 100% recall, 96% F1-score
Blue Tit: 98% precision, 92% recall, 95% F1-score
Robin: 98% precision, 92% recall, 95% F1-score
Great Tit: 92% precision, 97% recall, 94% F1-score
Chaffinch: 92% precision, 94% recall, 93% F1-score
House Sparrow: 98% precision, 94% recall, 96% F1-score
Blackbird: 98% precision, 95% recall, 96% F1-score

Error Analysis: Inter-Class Similarity

Confusion matrix analysis shows that classification errors mainly occur between visually similar species:

Blue Tit vs Great Tit: Both tits have similar color patterns and body shapes
House Sparrow vs Dunnock: Similar size and brown tones make distinction harder
Robin variations: Differences within species due to variations in color intensity and body shape at different ages

UI/UX Design: Figma Prototype

User-Centered Design Approach

Complete app interface designed in Figma with focus on intuitive navigation and clean visual hierarchy:

📸 Camera & Upload

Direct camera access for instant photos or upload from gallery for identification

🗺️ Spotting Hotspots

Map with popular birdwatching locations for community engagement

📚 Species Directory

Overview of most common species with search functionality

ℹ️ Detail Pages

Extensive information per bird species with characteristics and habitat info

A/B Testing: Color Scheme Optimization

🎨 User Testing Insights

An A/B test with 10 users showed a clear preference for the more colorful variant. The vibrant colors resonated better with the nature theme and made the app visually more appealing to the target audience of birdwatchers and nature enthusiasts.

Design Deliverables

Wireframes: Initial low-fidelity sketches to validate user flow
High-fidelity mockups: Detailed designs with final color scheme and typography
Interactive prototype: Clickable Figma prototype for user testing
A/B test variants: Two complete color scheme versions tested with target audience

Prototype Walkthrough

Technology Stack

🧠 MobileNetV2

🔥 TensorFlow/Keras

🐍 Python

🎨 Figma

📊 NumPy/Pandas

🖼️ Scikit-Image

📈 Matplotlib

🔍 Scikit-Learn

Complete ML Workflow

Data collection: Google Image Scraper for automated image gathering
Data preprocessing: Scikit-Image for resizing, MinMaxScaler for normalization
Model development: TensorFlow/Keras for CNN architecture and transfer learning
Training pipeline: ImageDataGenerator for augmentation, Adam optimizer, early stopping
Model evaluation: Scikit-Learn for classification reports, confusion matrices
Visualization: Matplotlib for training curves and performance analysis
UI/UX: Figma for complete app interface design and prototyping

Technical Challenges & Learnings

Transfer learning is powerful for small datasets: MobileNet's pre-trained weights enabled 95% accuracy with only ~400 images per class.
Data augmentation is essential: Doubling dataset size via augmentation significantly improved model robustness without additional data collection.
Manual data curation matters: Investing time in carefully selecting scraped images prevents label noise and increases model quality.
Inter-class similarity requires attention: Similar bird species (like tits and sparrows) are naturally harder to distinguish and require more diverse training examples.
Early stopping prevents overfitting: Monitoring validation loss and stopping when no improvement saves training time and improves generalization.
User research validates design choices: A/B testing with real users gave concrete feedback on color preferences that would otherwise be purely speculative.

Presentation & Feedback

"Presentation to instructors was received positively. The combination of high model accuracy (95%) and a well-developed Figma prototype showed that both technical and design aspects were well executed."

What I Would Do Differently

Add more bird species to make the app more useful for Dutch birdwatchers.
Collect more diverse training data per species, especially for edge cases (different angles and lighting conditions).
Experiment with newer architectures, such as EfficientNet, for possibly better accuracy.
Develop an actual Flutter implementation to validate mobile deployment feasibility.