MobileNet-ImageNet Classifier

A more user-friendly demo version of the Android app that I developed for the experimental measurements of my 2018 master's thesis Performance of Neural Network Image Classification on Mobile CPU and GPU for Aalto University.

To reduce repository size and avoid possible license issues, framework libraries and neural net model files are not included here. The app can be modified to use basically any image classification model, but the demo uses MobileNet trained with 1000-class ImageNet dataset (see details below).

Neural Net frameworks:

Snapdragon NPE manually downloaded from Qualcomm website (requires registration). Note: the app was developed using SNPE version 1.10 (tested up to 1.13) so with newer versions might not work anymore

TensorFlow and TensorFlow Lite compiled at build time from Google's Maven repo

Image Classification models:

• TensorFlow-Slim -- untrained Inception V2 and MobileNet 1.0 were used in my thesis's measurements

• TensorFlow Lite models -- the demo app uses Mobilenet 1.0 224 Float

• Caffe-Mobilenet -- for compatibility reasons the demo app uses this for SNPE and TF, conversion back to TF was done with Caffe-TensorFlow

Screenshots

Demo Video