Mobile device platform for automated visual retail product recognition

Abstract

Disclosed is a mobile electronic device platform for automated visual product recognition. As part of the platform, a mobile electronic device comprising a processing element and computer-readable media performs automated product-recognition processes based on input image frames received from a photographic element. The non-transitory computer-readable media may include computer-readable instructions stored thereon, the instructions instructing the processing element of the mobile electronic device to complete the following data processing steps without sending or receiving the processed data over an active data connection to any other computing device: (1) pass an input image frame of the input image frames depicting a product through a convolutional neural network stored on the computer-readable media; (2) generate a product classification for the product based at least in part on passage of the image frame through the convolutional neural network; and (3) record product metadata corresponding to the product to a record of the automated product-recognition process. The convolutional neural network may be trained through adjustment of model parameters according to the output of multiple classification arms sharing upstream convolutional features and utilize network compression and/or quantization techniques described herein, resulting in significant performance gains.