Abstract:
Social media platforms are not just a source of communicating with each
other but also the main source of generating web content like images. Users
upload the images along with the tags describing the content of the image
or the context in which the image has been uploaded. But these user pro vided tags are not enough for successfully retrieving this web content due to
incorrect or noisy tags . To tackle this problem of incorrect or noisy tags
and in some cases the absence of tags, we develop a pipeline to assign tags to
images by using image content, provided tags and social context associated
with the images.
In out framework we propose a deep architecture for assigning new tags to
the images by using Darknet based pre-trained YOLO V3 find image content,
ResNet-18 to fetch the image features and two layered neural network based
Word2Vec’s skipgram to perform the task of tags generation.
We use YFCC100M dataset for performing experiments. Results shows that
our model performs well in performing the task of tag recommendation on
YFCC100M dataset and has shown a 46% improved accuracy as compared
to already available tags.
