# Capstone Project Presentation --- ## Fruit Classification <br> Student: Nguyen Anh Tu Instructors: Huy Le, Ha V.N, Tung Cao, Hung Ngo Course: Machine Learning for AI 2022 Le Hong Phong Highschool for the Gifted, HCMC ---  For project demo, please visit: https://huggingface.co/spaces/Devaholic/fruit-demo --- ## Task description $T$ - <p class="fragment">Classify fruits.</p> - <p class="fragment">Input: An image of 1 type of fruit.</p> - <p class="fragment">Output: Predicted label.</p> --- ### Quick walk-through the dataset (Experience $E$) <br> - [Fruits 360 Dataset](https://www.kaggle.com/datasets/moltean/fruits) on Kaggle. - Over 90000 images of more than 100 types of fruits and vegetables. - Each image is an image of **ONE** type of fruits and vegetables. --- ### Model Architecture <br> - Inception Architecture with InceptionV3 pre-trained model. - Fully connected layer with 512 units. - Output layer with 131 units and Softmax activation for 131 different classes. --- ### Inception Architecture <br> <div class="r-stack"> <img class="fragment fade-out" data-fragment-index="0" src="https://i.imgur.com/njDb6Ny.png" width="600" /> <img class="fragment current-visible" data-fragment-index="0" src="https://i.imgur.com/Nd8oQx4.png" width="600" /> </div> --- ### Accuracy & Loss |Models|Validation Accuracy|Validation Loss| |------|-------------------|---------------| |No pre-trained model used|0.9781|0.1177| |Model 1 with InceptionV3|0.9814|0.0842| |Model 2 with InceptionV3|0.9853|0.0720| |Model 3 with InceptionV3|0.9940|0.0318| --- <div style="display: flex"> <image src="https://i.imgur.com/K7kMWqE.png"></image> <image src="https://i.imgur.com/Zltv5xj.png"></image> </div> --- ### Algorithm for training <br> - Train with 100 epochs. - Use EarlyStopping to optimize training time. --- ## Conclusion - High accuracy score but cannot generalize well. - Main objectives is to improve the generalization of the model. --- ## Lesson learned - <p class="fragment">How to apply pre-trained model.</p> - <p class="fragment">Pre-processing images</p> - <p class="fragment">How to design a project structure.</p> --- ## Further work for improvements - Automatically (Web scraping)/ Manually collect more data. - Find other datasets. - Try tuning hyper-parameters. - Try other pre-trained models. --- # Thank you for listening