• Objective: Develop an image classification ML model for diverse categories such as sunsets, mountains, and beaches.
• Functionality: Categorize and identify images with high accuracy.
• Added Feature: Offer travel recommendations based on user preferences.
• Purpose: Address the need for a dependable image classifier in travel planning.
• Impact: Enhance the overall travel planning experience.

Peers’ Comments:

- Model Processing: Curiosity about the choice of CNN for image classification, and the metrics for gauging user engagement and model performance.
- Recommendation: The integration of the image classification and recommendation systems, addressing biases.
- Data Scrapping: Ensuring data quality from user-generated content on Flickr. The specific step takens for data handling and preprocessing.

Professor’s Feedback:

- Project’s Scale: Suggestion on Scaling down the project from integration of other systems.
- Recommendation Feedback: Inquiry on the validity and possibility of evaluation for a Recommendation Model.

Flickr API Consideration:

- Authentication: Setting up the Flickr API key using setFlickrAPIKey
- Rate Limiting: Understand and handle rate limits to ensure uninterrupted data scraping.
- Ethical Considerations: Consider privacy implications and attribution requirements.

Challenges of Data Scraping From Flickr:

- Conversion between API Calls and Images Downloaded: Data returns a list of characters with information of ID, URL, TAGS etc for the given image, instead of actual images.
- Data Duplication: API calls from Flickr return multiple duplications of images due to unknown backend functions.
- User-Generated Data: Inherent variability, potential biases, and noise, LOTS AND LOTS OF NOISES!!!!!.

1. Set Up Environment: Load the necessary libraries (FlickrAPI, dplyr, readr, httr) and set Flickr API key using setFlickrAPIKey().
2. API Call Function: Create a function to fetch photos with given tags, using pagination to avoid duplicates and limit tag words to reduce spam.
3. Filter and Select: From the fetched photos, select relevant columns and filter out unwanted tags to avoid specific words.
4. Download Images: Execute the download_images function with the filtered URLs to download and save images to a specified folder on your local machine.

1. Download 7,000 images from Drive, allocating them into ‘train,’ ‘validation,’ and ‘test’ folders, each with ‘city,’ ‘mountain,’ and ‘river’ subfolders.
2. Distribute images into these subfolders: 600 for training, 200 for validation, and 100 for testing in each category.
3. There will be a total of 900 images per category divided amongst the training, validation, and testing folders.
4. Implement VGC 16 application functions extract_features() and reshape_features() for data preprocessing.
5. Construct a 7-layer CNN model with 2,138,755 parameters and fit the processed data into this model for training and validation.

1. Utilized the pre-trained VGG16 model to create features and defined a densely connected classifier for data processing.
2. Analysis of performance graphs indicates that the VGG16-based model (model2) demonstrates superior performance.
3. The performance of model2 on both training and validation tests is satisfactory for finalization.

1. The VGG16 improves the accuracy to ~82% which is considered to be good for image classification model.
2. The better performance of fine-tuning the VGG16 and training a CNN from scratch is visible.
3. Due to the quality and quantity of data, we get varied results for this project and the model slightly overfits the training data.

## Model: "sequential_1"
## ________________________________________________________________________________
##  Layer (type)                       Output Shape                    Param #     
## ================================================================================
##  dense_6 (Dense)                    (None, 256)                     2097408     
##  dropout_4 (Dropout)                (None, 256)                     0           
##  dense_5 (Dense)                    (None, 128)                     32896       
##  dropout_3 (Dropout)                (None, 128)                     0           
##  dense_4 (Dense)                    (None, 64)                      8256        
##  dropout_2 (Dropout)                (None, 64)                      0           
##  dense_3 (Dense)                    (None, 3)                       195         
## ================================================================================
## Total params: 2138755 (8.16 MB)
## Trainable params: 2138755 (8.16 MB)
## Non-trainable params: 0 (0.00 Byte)
## ________________________________________________________________________________

## 10/10 - 0s - loss: 0.5228 - accuracy: 0.7767 - 126ms/epoch - 13ms/step

##      loss  accuracy 
## 0.5227583 0.7766666