This test uses a pre-trained neural network model to recognize objects in a set of photographic images. This technique has many practical uses in mobile applications such as identifying and classifying images in photography apps, tagging people and places in social networking apps, and helping visually impaired people understand the world around them.
A full explanation of machine learning, image recognition techniques, and neural networks is beyond the scope of this guide. We recommend following the links in the text below for further reading.
TensorFlow is an open-source machine learning library developed and supported by Google. "Inception" is a deep convolutional neural network architecture developed by Google for TensorFlow that is part of the current state of the art for computer image recognition.
The benchmark test uses a TensorFlow Inception model that has been pre-trained with the ImageNet database. ImageNet is an academic data set containing thousands of images that is commonly used for training image recognition systems.
The Inception model is about 52 MB in size, making it suitable for mobile devices. The model is loaded into memory and tested with a set of 10 images. Each image is reduced to 256 × 256 vector size before testing.
For each image, the model returns a percentage that represents its confidence that the image was recognized successfully. These percentages are not used in the scoring, however. Since the same model is used on all devices, the confidence of recognizing each image is also the same on all devices.
Instead, the test measures how long it takes the model to classify each image. The time taken depends on the performance of the processor in the device.
The result is the average of the times taken to recognize each of the 10 images.