Bird vs Drone Detection and Classification

Aim : To develop a deep learning-based solution that can classify aerial images into two categories Bird or Drone and optionally perform object detection (YOLO) to locate and label these objects in real-world scenes.

Project Takeaways:

• Deep Learning
• Computer Vision
• Image Classification & Object Detection
• Python
• TensorFlow/Keras
• Data Preprocessing & Augmentation
• YOLOv8
• Model Evaluation
• Streamlit Deployment

WorkFlow

- Image Classification Training (Custom and Transfer Learning Model) Aerial_Object_Detection.ipynb

• Used cv2 to show an example on Data augmentation

• Used ImageDataGenerator library to create augmented data in training_data

  • Resize, rotation, zoom, horizontal view, vertical view

• Data Preprocessing

  • Resize images to (224,224) pixels

• Custom Training from Scratch

• Introducing Transfer Learning with 3 different pre-trained applications from tensorflow.keras.applications

• Model Evaluation

• Chose the best model among the 4 and saved the model for future use.

  

- Image Detection Training (YOLO Model) YoloModel.ipynb

• Data Preperation
  • Yolo requires the training images to be in an particular folder dir with train,validation and test data split in labels and images folders seperately.
• Creating data.yaml file for model training data.yaml
• Model Training
• Model Evaluation

• Saving Model

- Streamlit Application streamlit.py

• Side bar to upload data and select a model to view result.
• In main content the left side shows original image and right side with the predicted output.
• Created an download button to download YOLO detected image

- Image Classification

- Image detection