Object identification for inventory management using convolutional neural network

In this paper, a real time inventory monitoring system using Convolutional Neural Network (CNN) has been proposed. The objective is to develop a reliable and efficient method for object counting and localization in the inventory using vision interface. Firstly, the Region of interest (ROI) is extracted from the image of inventory using visual features i.e. stable regions of input image. The centroids for these stable regions are estimated using Connected Component Analysis (CCA). These centroids are used to get bounding box in the stable regions. The bounding box generated for their respective ROIs are fed to CNN followed by Softmax layer for classification. Herein, single layer CNN architecture is used. The Softmax layer is used to classify the bounding boxes and final box is drawn for approximate location of object with distance, measured using RGB-D sensor. The CNN training has been performed using STL-10 Dataset. The Softmax layer is trained for various classes with their respective labels. The test has been performed for two different classes. The developed algorithm performed well irrespective of resolution and color variations in input image and able to count objects with respect to depth parameter.

I. Introduction

Manual inventory management results higher inventory carrying cost due to errors in accounting of objects. It is a labor intensive and time consuming process which requires continuous monitoring of appropriate stock level present in inventory. Automatic inventory management using vision interface is important aspect in industrial applications such as objects identification and counting [1]. It comprises of autonomous system which provides count of each object present in inventory in order to reduce the inventory carrying cost. Automated inventory management [2] increases the efficiency with high precision or accuracy over the inadequacy of manual inventory management.