深度学习在泰文商品识别中的应用

Application of Deep Learning in Thai Commodity Recognition

跨境零售是中泰贸易的关键,但传统采用RFID标签识别和人工问询模式的购物体验度并不高。因而运用基于机器视觉的商品识别方法,构建"互联网+跨境贸易"的人工智能化零售电商体系,对推动中泰"一带一路"的区域经济发展具有重要的应用价值。该文在图像分类识别技术和卷积神经网络研究的基础上,提出一种基于Inception-V3框架参数迁移学习的Inception-Thai神经网络模型来用于泰文商品的识别,可避免在少量样本上出现训练过拟合的现象。在人工标注的7类泰文商品图片数据集上进行测试,实验结果表明,该模型具有较高的图片深层特征提取能力,且在商品识别的分类置信度上可达到83%至98%,取得了较高的准确率。

Cross-border retail is the key to Sino-Thai trade, but the traditional shopping experience using RFID tag recognition and manual inquiry mode is not satisfactory. Therefore, the use of machine vision-based commodity identification method to build an artificial intelligence retail e-commerce system of "Internet + cross-border trade" has important application value for promoting the regional economic development of the China-Thailand "Belt and Road". Based on the research of image classification and recognition technology and convolution neural network, this paper proposes an Inception-Thai neural network model on the basis of Inception-V3 framework parameter transfer learning for Thai commodity recognition, which can avoid over-fitting in a small number of samples. The experimental results on seven types of Thai merchandise image datasets show that the model has a high ability to extract deep-seated features of images, and the classification confidence of merchandise recognition can reach 83% to 98%, which achieves a high accuracy.