Systems and methods for improving visual search using summarization feature

A computer-implemented method for training a metric learning convolutional neural network (CNN)-based model for a target cross-domain image retrieval application, the computer-implemented method executable by a processor, the method comprising:

• generating a plurality of batches sampled from a cross-domain training dataset to train the metric learning CNN-based model to match one or more images of a first domain with one or more images of a second domain,
• wherein each image of the one or more images of the first domain and the one or more images of the second domain belong to one of a plurality of sub-categories, each sub-category belonging to one of a plurality of categories,
• wherein each batch of the plurality of batches comprises a first plurality of images sampled from a first number of sub-categories,
• wherein each sub-category of the first number of sub-categories of each batch of the plurality of batches comprises a first set of images sampled from the first domain, and a second set of images sampled from the second domain,
• wherein the first set of images comprises a second number of images, and
• wherein the second set of images comprises a third number of images;
• and
• training the metric learning CNN-based model using the generated plurality of batches,
• wherein the metric learning CNN-based model generates an embedding vector from a given input image,
• wherein the metric learning CNN-based model comprises:
• a first pooling layer pooling an output feature map of a first convolutional layer, generating a first pooled feature vector;
• a first normalization layer normalizing the first pooled feature vector, generating a first normalized feature vector;
• a second pooling layer pooling an output feature map of a second convolutional layer different from the first convolutional layer, generating a second pooled feature vector;
• a second normalization layer normalizing the second pooled feature vector of the second pooling layer, generating a second normalized feature vector;
• a concatenation layer concatenating the first normalized feature vector with the second normalized feature vector, generating a concatenated vector;
• and
• one or more fully connected (FC) layers reducing a dimensionality of the concatenated vector to a lower output embedding vector dimensionality,
• wherein the using of the generated plurality of batches comprises executing a loss function based on one or more batches of the plurality of batches,
• wherein the loss function is selected from the group consisting of a triplet loss function, a contrastive loss function, and a cluster loss function,
• wherein the loss function is configured to allow the metric learning CNN-based model to learn a mapping from the cross-domain training dataset to an embedding space,
• wherein a similarity between a first given image and a second given image of the cross-domain training dataset corresponds to an embedding distance between the first given image and the second given image, and
• wherein the embedding distance between the first given image and the second given image is a distance between the corresponding embedding vectors generated from the first given image and the second given image using the metric learning CNN-based model.