Skip to content

Deep Learning based Waste Segregation Project to classify waste images into different classes.

License

Notifications You must be signed in to change notification settings

vasantvohra/TrashNet

Repository files navigation

TrashNet Build Status Udacity - Intro To Tensorflow

Linkedin Article

Waste Segregation Project to classify waste into different classes.

Already deployed version

DIRECT LINK

Kaggle Kernel Dataset: Trashnet

Categories
Cardboard
Glass
Paper
Metal
Trash

Just a Beginner!

First attempt: Building Tensorflow keras CNN Model

Notebook

  • Training on GrayScale images
  • Validation Accuracy 42%
  • Loss function : Sparse Categorical Loss function
  • Overfitting High

Second attempt:

Notebook

  • Image Augmentation
  • Training on Colored images
  • Validation Accuracy 80%
  • Loss function : Categorical Loss function
  • Added 1 more 32 filters Convolution block with default stride
  • 2 Dense layers with dropouts

Understood

  • Image, Fit, Predict Generators, Flow from directory.
  • Difference between SpatialDropout2D and Dropout Regularization.
  • Number of filters and dense perceptrons to build a model.
  • Callbacks : Early Stopping and Model Checkpoints to save perfect model on Hierarchical Data Format HDF5 (.h5)
  • Visualizations by Matplotlib

Completion of Course on :

Udacity Intro To Tensorflow
Completion

Next Steps

  1. Further Regularization -
    • Batch Normalization
    • L1 & L2 error
    • intialization of weights
  2. Transfer learning -
    • MobileNet
      • Saving & Deploying of TFLite
    • VGGNet
    • ResNet
  3. Collect Preprocess my own Training Dataset.
  4. Object Detection Localization
    • YOLO v2/v3
  5. Trash, Instance Segmentation

Understood Git Large File Storage

Large H5 Files