Object classification using CNN & VGG16 Model (Keras and Tensorflow)
- 1. Object Classification using CNN Lalit Jain https://www.linkedin.com/in/lalit7jain/ Rohit Agarwal https://www.linkedin.com/in/rohitag04/ Shivam Goel https://www.linkedin.com/in/goelshi/ Hina Gandhi: https://www.linkedin.com/in/hina-gandhi-52834356/ Big Data & Engineering Systems Under the guidance of Professor. Dino Konstantopoulos
- 2. Agenda 1. Case Study 2. Building a Convolutional Neural Network 3. Time Complexity 4. Using VGG16 model 5. Demo 6. Results 7. Limitations & Challenges 8. Future Scope Agenda
- 3. Case Study Building a CNN model which can be trained on the fly and classify objects
- 4. Approach 1. Using Google API to download 100 images from the front end using a Keyword 2. Prepare the image by cleaning and augmentation 3. Train the model based on the image downloaded with Convolutional Neural Network using Keras and Tensorflow 4. Classify the image Technology Stack: Programming Language: Python Front End: Django Framework Machine Specifications: Libraries: Keras, Tensorflow, VGG16, etc
- 6. Images are a matrix of pixel values Convolution Operation Relu Activation Pooling (Max)
- 7. Putting it all together
- 8. About the Dataset We have taken two datasets : Static Model: One dataset is related to natural calamities where we have trained our models for 4 classes “earthquake”, “hurricane”, “tornado” and “volcanic eruption”. Dynamic Model: Other dataset is related to personality identification which will work on the fly. During demo, we will add one personality on the fly in addition to the existing classes (Donald Trump, Morgan Freeman)
- 9. Use case and Scope 1. Using static model, identify one of the 4 natural calamities 2. Using the trained dynamic model, identify the personality Out of Scope/ Limitations: 1. Being able to identify a class which is not trained 2. Being able to identify 0 as output if image does not match 3. Misclassification of images due to training on uncleaned and inaccurate image labels
- 10. DEMO
- 11. Data augmentation for improving the model By applying random transformation to our train set, we artificially enhance our dataset with new unseen images. shear_range: Rotate image by 0.2 zoom_range: Zoom the image by 0.2 horizontal_flip: Randomly flip inputs horizontally rescale: rescaling factor (1./255)
- 12. Building CNN from Scratch Basic CNN model (the model that was developed from scratch) gave ~74% accuracy.
- 13. CNN training complexity Time: 8-9 Hours Processor: CPU Dataset: 2000 images Classes: 2 Can we improve it?
- 14. VGG16 Pre-Trained Model https://arxiv.org/pdf/1409.1556.pdfCitation: The model achieves 92.7% top-5 test accuracy in ImageNet, which is a dataset of over 14 million images belonging to 1000 classes.
- 15. Using the VGG16 Model
- 16. Using the VGG16 Model We have created our own sequential model and then added to VGG16.
- 17. Improvement Using this approach, the accuracy was improved to 87%. Time to run the model was still 2-3 hours. Can we improve the time?
- 18. Using GPU machine Running the model on a AWS instance with Nvidia GK104 GPU (4 GB) powered instance The epochs that took 1 hour to train 140 images are now executing in under a minute as shown in next slide with an accuracy of 95%
- 19. Results
- 20. Results Evolution of accuracy on training (blue) and validation (green) sets for 1 to 30 epochs :
- 21. CHALLENGES 1. Limitation of Google API custom search to return only 10 images. 2. Compressing large size images before writing to disk. 3. Basic CNN model took more than 8 hours to run from scratch 4. The total images in all the classes should be same and the batch size should be the GCD of total images in train and test classes. 5. Django installation and other software installations required lot of configurations in GPU instance 6. Tried to install quiver engine for visualization of each convoluted layer.
- 22. Future Scope 1. Facial key point detection 2. Analysis of satellite images for disaster detection 3. Real Time Criminal Detection through Video analysis 4. And many more…