基于卷积神经网络的商品图像识别

Recognition of Commodities Images Based on the Convolutional Neural Network

在无人零售与智能超市等新型商业终端环节中需要准确识别商品的种类等信息,但实际识别过程中往往会受到复杂背景、光照不均、角度距离等多种因素的干扰,对识别算法的泛化能力提出了更高要求。为解决上述问题,本研究利用深度学习中的卷积神经网络(CNN)模型对商品图像进行了系列识别研究,通过模型的自主学习特性和强泛化能力,实现复杂条件下的商品图像识别,主要内容包括:首先,以烟包图像为例,分析了商品图像识别在实际应用中存在的具体困难,通过改变环境背景、光照条件、拍摄距离、商品角度等条件采集了系列商品图像信息样本;同时,构建了双层卷积神经网络模型,在完成图像处理后开展了卷积与池化计算,获取了商品图像的多层次抽象化特征;进而,对获取特征集进行了全链接并集成了分类器,实现了复杂环境下商品图像的分类,对中间特征图像进行了抽取与分析;最后,将同样数据集输入BP神经网络、RBF神经网络、SVM等进行了对比,在2组对比实验中最高识别准确率分别为90.48%和78.48%,明显低于CNN模型的98.42%和98.52%。所提出方法可有效克服常见场景的因素干扰,在无人零售、智能超市、商品检测等领域具有广泛应用价值。

In the new business terminal links such as unmanned retail and intelligent supermarket,the types of goods and other information are needed to be recognized.However,in the actual recognition process,some strict requirements are put forward for the recognition algorithm due to various interference conditions such as complex background,uneven illumination,angle distance and others.In order to solve the problems above,a series of commodity image was recognized based on the convolution neural network in this work.Recognition for commodity images under complex conditions was realized with the self-learning characteristics and strong generalization abilities in deep learning.The main contents were as follow.Firstly,common cigarette packaging images were taken as recognition objects,and a series of samples were collected by changing the environmental background,lighting conditions,shooting distance,commodity angle and other conditions.Secondly,a two-layer convolution neural network model was constructed to realize the calculation of convolution and pooling,multi-level abstract features of commodity image were obtained.Then,the feature set was fully linked and a classifier was integrated to realize the classification of commodity images in complex environment,the intermediate feature images were also extracted and analyzed.Finally,the same data set was also input into BP neural network,RBF neural network and support vector machine.In these two groups of experiments,the highest recognition accuracy was 90.48%and 78.48%,which was obviously lower than 98.42%and 98.52%in the CNN model.The proposed method can effectively overcome the interference of common scenes,and has wide application value in many fields such as unmanned retail,intelligent supermarket and commodity detection.