Transfer Learning Neural Network implementation for food recognition
- 1. Estimating Calorie Content in Food Dishes using Transfer Learning Neural Networks Ali Abul Hawa
- 2. 2
- 3. Difficulty in this project/dataset: many food dishes look very similar Beef noodleRamen noodle 3
- 4. 4 Low-level features Higher-level features Output Classification Layers 1.2M images 1000 classes 16 Layers VGG16, 16 layers: 3 weeks on 4 GPUs Titan Black ~$6000 My computer (~ g2.2xlarge AWS) : ~ 30 weeks TRANSFER LEARNING
- 5. 5 Low-level features Higher-level features Custom Output Classification Layers 15 Layers 100 Food classifier Train custom classification layer ~52% TRANSFER LEARNING
- 6. 6 Low-level features Higher-level features Custom Output Classification Layers 15 Layers 100 Food classifier Train classification layer + highest level feature layer ~ 70% top-1, 92% top-5 accuracy in 30 hours 竕ｪ 30 weeks TRANSFER LEARNING
- 7. TRANSFER LEARNING Why transfer learning? Time Works well even on small datasets. However, need to use image augmentations to prevent overfitting! 7
- 8. 8 Live Demo 窶ｦ 窶ｦ Thanks!
Editor's Notes
- #3: Summary: multiclass classification of food images using transfer learning deep neural networks. From an image predict the food inside it, and return the calorie content per serving Why? I chose to work with neural networks because they can do fascinating work, and I believe thy are amazing. Also, I wanted to do something challenging窶ｦ I窶冦 still learning a lot about them, and there is much more to learn. Why food prediction? It窶冱 a nice idea. Most people want a healthy lifestyle, including food habits, the way to do that is by knowing at least the calories before consuming them. When I mentioned this to some ladies, they were excited about it, and they wanted the product now. Importance: this can used with different applications, and uses: for example, it can be an added End game: classify images into 1 of 100 classes窶ｦ easy, right? No 窶ｦ窶ｦ.explain Dataset description: Number of classes 100 Number of images 14600 # images/class ~ 150 Size of images varies from 80x140 -> 800x800
- #5: Main Idea of transfer learning: use a NN that has been pre-trained intensively on a huge dataset for any class/types. Use the super capabilities of this NN in extracting image features, such as lines, edges, blobs, shapes, 窶ｦetc Remove the last layer (classifier) designed for a particular image classification problem, such as faces/animals/cars. Use this NN as a feature extraction device connected in between your images and your custom NN (classifier) Further improvement by fine-tune training the last layer of the super-pre-trained NN. -> quick + accurate