Samsung Sds America, Inc.

Hankyu Moon

Sima Didari

Heng Hao

Jae Oh Woo

Patrick David Bangert

Training of a machine vision model, a segmentation model, is performed by using an acquisition function for a small number of pixels of one or more training images. The acquisition function uses first mutual information and second mutual information to identify unlabelled pixels which are labelled with high uncertainty when predicting possible label values. Training, prediction of labels, identifying pixels with highly uncertain labels, obtaining labels only for those pixels with highly uncertain labels and retraining are performed iteratively to finally provide the machine vision model. The iterative approach uses very few labelled pixels to obtain the final machine vision model. The machine vision model accurately labels areas of a data image.

Bayesian semantic segmentation active learning with beta approximation

A problem of supervised learning is overcome by using patches to discover objects in unlabeled training images. The discovered objects are embedded in a pattern space. An AI machine replaces manual entry steps of training with a machine-centric process including clustering in a pixel space, clustering in latent space and building the pattern space based on different losses derived from pixel space clustering and latent space clustering. A distance structure in the pattern space captures the co-occurrence of patterns due to frequently appearing objects in training image data. Embodiments provide image representation based on local image patch naturally handles the position and scale invariance property that is important to effective object detection. Embodiments successfully identifies frequent objects such as human faces, human bodies, animals, or vehicles from unorganized data images based on a small quantity of training images.

Object discovery using an autoencoder

A problem of imbalanced big data is solved by decoupling a classifier into a neural network for generation of representation vectors and into a classification model for operating on the representation vectors. The neural network and the classification model act as a mapper classifier. The neural network is trained with an unsupervised algorithm and the classification model is trained with a supervised active learning loop. An acquisition function is used in the supervised active learning loop to speed arrival at an accurate classification performance, improving data efficiency. The accuracy of the hybrid classifier is similar to or exceeds the accuracy of comparative classifiers in all aspects. In some embodiments, big data includes an imbalance of more than 10:1 in image classes. The hybrid classifier reduces labor and improves efficiency needed to arrive at an accurate classification performance, and improves recognition of previously-unrecognized images.

Growing community of inventors

San Ramon, CA, United States of America

Hankyu Moon