AI/ML methods to retrieve and classify images related to Climate Events -- 2
Budget: £250 – £750 GBP
This project aims to apply some suitable AI/ML methods to search the internet, the images in particular, and other multi-media materials, such as videos if possible, to retrieve the truly relevant climate events, and to extract and classify them into some predefined categories.
1. Retrieve images from internet - build some custom search engine algorithm like google (https://developers.google.com/custom-search/docs/queries)
you can put these terms as “Flood AND UK AND August 2018” to Google Image search or use its API
2. form Training dataset and testing dataset from retrieved images (make sure all the retrieved images are relevant ones. delete unrelated images manually if necessary.
3. build image classification model where the images which we retrieved from internet should be predicted as truly relevant climate events and then classify the truly predicted images into folders.
more notes useful for ideas and approaches to do this project:
1. Try to find and/or collect as many labelled images as possible for each climate type event from known images database, and use them in a usual way to train CNN models and test their performance
pass 3 parameters 1. climate event name 2. location where it occurred 3. which year the event occurred.
then download the images,
2. If these CNN models are good enough, you can use them to classify the retrieved images
3. Manually check some samples of classified images to validate their relevance.
4. For the retrieved images, you should try to extract their meta data out, which may contain some information about the event captured by the images in their caption, time and location. You can then use these pieces of information with some text analysis to verify the classified images.
1. Retrieve images from internet - build some custom search engine algorithm like google (https://developers.google.com/custom-search/docs/queries)
you can put these terms as “Flood AND UK AND August 2018” to Google Image search or use its API
2. form Training dataset and testing dataset from retrieved images (make sure all the retrieved images are relevant ones. delete unrelated images manually if necessary.
3. build image classification model where the images which we retrieved from internet should be predicted as truly relevant climate events and then classify the truly predicted images into folders.
more notes useful for ideas and approaches to do this project:
1. Try to find and/or collect as many labelled images as possible for each climate type event from known images database, and use them in a usual way to train CNN models and test their performance
pass 3 parameters 1. climate event name 2. location where it occurred 3. which year the event occurred.
then download the images,
2. If these CNN models are good enough, you can use them to classify the retrieved images
3. Manually check some samples of classified images to validate their relevance.
4. For the retrieved images, you should try to extract their meta data out, which may contain some information about the event captured by the images in their caption, time and location. You can then use these pieces of information with some text analysis to verify the classified images.