Personalized Visual Recognition via Wearables: A First Step Toward Personal Perception Enhancement
| Autor: | Hosub Lee, Steven Eliuk, Alfred Kobsa, Cameron Upright |
|---|---|
| Rok vydání: | 2017 |
| Předmět: |
Multimedia
business.industry media_common.quotation_subject Deep learning Cognitive neuroscience of visual object recognition Reverse image search Wearable computer 020207 software engineering 02 engineering and technology computer.software_genre Personalization Task (project management) 020204 information systems Perception 0202 electrical engineering electronic engineering information engineering Artificial intelligence business Psychology Transfer of learning computer media_common |
| Zdroj: | Intelligent Systems Reference Library ISBN: 9783319625294 Personal Assistants |
| DOI: | 10.1007/978-3-319-62530-0_6 |
| Popis: | During the last few years, deep learning has led to an astonishing advancement in visual recognition. Computers now reach near-human accuracy in visually recognizing characters, physical objects and human faces. This will certainly allow us to build more intelligent personal assistants that can help users better understand their surrounding environments. However, most visual recognition systems have been designed for user-independent recognition (e.g., Google reverse image search), and not for an individual user. We believe this practice is restricting the technology from helping people who have individual needs. For example, a person with memory problems may want to have a computer that accurately recognizes a few close friends, rather than hundreds of celebrities. To address this issue, we propose a novel wearable system that enables users to create their own visual recognition system with minimal effort. A client running on Google Glass collects images of objects a user is interested in, and sends them to the server with a request for a specific machine learning task: training or classification. The server performs deep learning according to the request and returns the result to Glass. Regarding the training task, our system not only aims to build deep learning models with user generated image data, but also to update the models whenever new data is added by the user. Experiments show that our system is able to train the custom deep learning models in an efficient manner, in terms of the required amount of computing power and training data. Based on the customized deep learning model, the system classifies an image into one of 10 different user-defined categories with 97% accuracy. |
| Databáze: | OpenAIRE |
| Externí odkaz: |
|