Abstract:
In this study, we propose a system that classifies streams as supervised learning and
recognizes brands as a one-shot learning problem. We also created a custom dataset for
implementing stream classification. This dataset comprises streams from multiple sources
with respective timestamps. For classification, we are using pre-trained ResNet on
ImageNet-1K dataset. Then the network is fine-tuned for the subject dataset, by keeping
ResNet feature extraction layers frozen. ResNet’s feature extraction layer is utilized for
vectorized embedding extraction for brand recognition. The final fine-tuned classification
layer is used for stream classification. By multiple experimentation of different networks, we
were able to achieve 83% accuracy on the stream classification. For brand recognition, we
were able to squeeze similar results on a benchmark dataset. Both these classifications can be
used to track temporal occurrence and length, using the same network to generate stream
analytics. Such a combination of supervised and one-shot learning can be helpful in other
applications as well.
