A CNN-based supermarket auto-counting system

Abstract

Deep learning has made significant breakthrough in the past decade. In certain application domain, its detection accuracy has surpassed human being in the same task, e.g., voice recognition and object detection. Various novel applications has been developed and achieved good performance by leveraging the latest advances in deep learning. In this paper, we propose to utilize deep learning based technique, specifically, Convolutional Neural Network (CNN), to develop an auto-counting system for supermarket scenario. Given a picture, the system can automatically detect the specified categories of goods (e.g., Head & Shoulders bottles) and their respective numbers. To improve detection accuracy of the system, we propose to combine hard example mining and multi-scale feature extraction to the Faster R-CNN framework. Experimental results demonstrate the efficacy of the proposed system. Specifically, our system achieves an mAP of 92.1%, which is better than the state-of-the-art, and the response time is about 250 ms per image, including all steps on a GTX 1080 GPU.