Create a tool to aid in filtering out misclassified images.
To find the images that are highly related to a set of query words:
1) Obtain a pretrained model for classification (probably trained on ImageNet)
2) Look up word vectors associated with the output categories and query words
3) Create a word vector corresponding to the set of query words (just average the word vectors)
4) Create a word vector corresponding to each image
Make category predictions on each image
Take a weighted sum of the word vectors for each category using the predicted probabilities as weights.
5) Find the images whose word vectors have the smallest cosine similarity with word vector for the set of query words.
Create a tool to aid in filtering out misclassified images.
To find the images that are highly related to a set of query words: 1) Obtain a pretrained model for classification (probably trained on ImageNet) 2) Look up word vectors associated with the output categories and query words 3) Create a word vector corresponding to the set of query words (just average the word vectors) 4) Create a word vector corresponding to each image