Indonesian Herbs and Spices Recognition using Smaller VGGNet-like Network

Abstract

This study proves that the use of Smaller VGGNet convolutional neural network, are very capable of performing image recognition tasks. In this study, the task was the recognition of herbs and spices images. The main difficulty of herbs and spices image recognition, including high object similarities and images that are usually taken in natural conditions without special preparation, both in terms of lighting and shooting angle. We built a Smaller VGGNet family with 1024 features and some additional dropout layer on the layer after the convolutional layer. However, in area of Indonesian herbs and spices, there is no such database publicly available to researchers, thus it is hard to evaluate different methods with the same standard. Therefore, image database in this study created by searching images in Google Image Search. After going through the selection stage, the image database has 3574 images for 27 classes. Result shows that our model is very capable of recognizing herbs and spices, with average labeling accuracy of 70%. We also found that adding dropout layer after convolutional layer, can help the model to reduce the overfitting, and indirectly improve system accuracy.