• A
  • A
  • A
  • ABC
  • ABC
  • ABC
  • А
  • А
  • А
  • А
  • А
Regular version of the site

Article

Многозначная классификация множества изображений на основе нейросетевого механизма внимания

Дёмочкин К. В., Савченко А. В.

In this paper we consider several efficient techniques for multi-label classification of a set of images. We propose the two-stage approach. First, we perform transfer learning on a pretrained convolutional neural network in order to use it as a feature extractor. Next, the feature vectors of each image from a given input set are combined into a single vector using the modification of the attention-based neural aggregation module. In the experimental study we examine the classification of the user's interests based on the photos of products purchased by this user using the Amazon Product Home and Kitchen dataset. It was shown that one of the most highest F1-measure (0.87 for 15 recommendations) is achieved for one-layer attention block with squeezed visual features. It is emphasized that the resulting model including the MobileNet feature extractor has 16 Mb size only.