基于卷积神经网络的物流货物图像分类研究
2021-08-21刘斌龙健宁程方毅龚德文
刘斌 龙健宁 程方毅 龚德文



摘要:基于卷积神经网络(CNN)针对物流环境下货物的图像分类问题进行了研究。首先,在实际物流环境下收集了13种货物的 ROI图像,并通过每隔10°旋转的方式来扩充数据集以防止过拟合现象的发生;然后,在考虑了实际硬件条件的情况下构建了轻量级 CNN ,并进行了基于自建数据集的训练,训练实验发现,轻量级 CNN模型具有很快的收敛速度并在验证集取得了100%的准确率;最后,研究了旋转对货物图像分类性能的影响,并进行了可视化分析,验证了 CNN对旋转操作基本不具备一致性。
关键词:卷积神经网络;图像分类;自动识别;旋转不变性
中图分类号:TP391.41 文献标志码:A 文章编号:1009-9492(2021)12-0079-04
Research on Goods Image Classification Based on Convolutional Neural Networkin Logistic Environment
Liu Bin1,Long Jianning1,Cheng Fangyi1,Gong Dewen2
(1. Key Laboratory of Polymer Processing Engineering of Ministry of Education//Guangdong Provincial Key Laboratory of Technique andEquipment for Macromolecular Advance Manufacturing// National Engineering Research Center of Novel Equipment for Polymer Processing,Guangzhou 510641, China;2. Guangdong Changheng Intelligent Technology Co., Ltd., Dongguan, Guangdong 523841, China)
Abstract: The classification of goods images in logistics environment based on convolutional neural networks has been studied. First, ROI images of 13 kinds of goods in logistics environment were collected and the data set was expanded by rotating every 10° to prevent overfitting. Then, a lightweight CNN was constructed and trained based on self-built data set considering the actual hardware conditions, and the training experiment showed that the model converged quickly and achieved 100% accuracy in the validation set. Finally, the effect of rotation on the classification performance of goods images was studied and performed a visual analysis, the result showed that CNN had poor invariance to rotation operation.
Key words: CNN; image classification; automatic recognition; rotation invariance
0 引言
近年來,随着电子商务的飞速发展,物流行业的业务量随之剧增,不断成熟的自动识别技术引入到现代物流当中以代替人眼对货物进行分辨识别,物流行业的自动化程度在不断提升的同时,物流过程的效率在很大程度上得以提升[1]。
目前,在物流行业中最为常见的两种自动识别技术为条形码技术以及 RFID 技术,但是,在诸如食品、医药以及化妆品等行业当中,不仅货物种类多样,而且其外表面没有或者难以张贴条码或射频卡等信息载体,这种情况下,就需要利用货物自身的形状、颜色以及表面纹理特征等,来实现货物种类的自动识别。传统的图像分类往往只能针对图像的某一特征,如颜色、轮廓、纹理等进行分类,因而鲁棒性和泛化能力不足,而鉴于物流环境下的复杂性,传统的图像分类就难以满足实际物流需求[2-3]。……
