Retail Product Classification on Distinct Distribution of Training and Evaluation Data

• Published in: Pattern recognition and image analysis Vol. 32; no. 1; pp. 142 - 152
• Main Authors: Jonathan, Kusuma, Gede Putra
• Format: Journal Article
• Language: English
• Published: Moscow Pleiades Publishing 01.03.2022; Springer Nature B.V
• Subjects: Accuracy; Application Problems; Artificial neural networks; Classification; Computer Science; Computer vision; Evaluation; Image classification; Image Processing and Computer Vision; Object recognition; Pattern Recognition; Training; deep learning; image classification; retail product classification; dilated convolution; convolutional neural network
• ISSN: 1054-6618; 1555-6212
• DOI: 10.1134/S105466182104012X

Retail product classification can be beneficial in the world of commerce, take for example helping vision-disabled parties in their shopping or evaluating product placement strategy. However, the available datasets for retail product classification are few and some have very distinct distribution of training and evaluation data, thus providing a huge challenge on its own. In addition, there are only few researches on this subject which can still be improved on. This paper attempts to improve previous approaches for retail product classification on very distinct training and evaluation data distribution by utilizing convolutional neural network (CNN) models inspired by well-performing CNN models in general image classification task, which later can be fine-tuned for other computer vision tasks, namely, object detection. The results show that VGG-16 performs at 66.9167% accuracy and a new modified VGG-16 model named VGG-16-D attains 66.83% accuracy with 85% fewer parameters than VGG-16, outperforming most existing approaches considering several comparison baselines.