Thanks to Artificial Intelligence (AI) & deep learning, visual recognition technology has made great strides of late, which has usually meant that your ever present mobile phone apps send images back to big back-end Cloud servers for further processing.
Now developers can add freely available image recognition models to their mobile apps & do their computing right on the phone without lag & without eating up the battery in the process.
Google’s deep learning, based on neural networks, & its Cloud Vision API has been one of the main players powering this revolution. It’s been enabling things like text recognition, but also object, logo & landmark recognition, though only for internet connected mobile devices.
Now MobileNets provides this functionality on-device without having to connect via the Internet to resource heavy server farms. Embedded apps can now take advantage of TensorFlow, Google’s open-source software library for machine intelligence, to get excellent accuracy in restricted resource environments.
Mobiles are low on space, power, memory & computational resources, so if you want to do image recognition locally you’ll need all the help you can get. Here, a number of parametrized small, low-latency, low-power models are provided so that an app designer can choose which one to use depending on their resource constrained user case.
The group of open source developers led by Google has made MobileNet v1 available on GitHub where it’s defined as:
…small, low-latency, low-power models parameterized to meet the resource constraints of a variety of use cases. They can be built upon for classification, detection, embeddings & segmentation similar to how other popular large scale models, such as Inception, are used. MobileNets can be run efficiently on mobile devices with TensorFlow Mobile.
There are 16 pre-trained models available, with Google having done all of the optimization for you ahead of time. The accuracy, latency & power usage scales with the number of multiply accumulates operations (MACs) you can spare on your processors. The models go from a low of 14 million MACs all the way up to 569 million, giving 66.2% top-5 accuracy on the low end & 89.5% on the high.
If you want to get started with MobileNets you might want to read this article from the Google Research Blog, & you´ll have to head on over to the TensorFlow-Slim Image Classification Library, also available on GitHub. Other resources on running the models on-device are also referenced on the Blog entry.
Image Credit: Google
•Share This•
