摘要:
本文以腊肉为实验对象,建立一种融合光谱曲线特征和图像特征的肉类食品分类与检索方法,利用10个3×3的卷积层、3个5×5的卷积层、5个5×5的池化层和2个全连层的CNN模型对高光谱图像进行特征提取,并以交叉熵作为优化目标,同时利用多元散射校正和主成分分析方法(Principal Component Analysis,PCA)对光谱特征进行预处理和特征提取,然后将两种特征进行融合,并将融合特征利用支持向量机(Support Vector Machine,SVM)进行分类。结果表明,直接使用CNN训练好的模型对高光谱图像进行特征提取,利用SVM作为分类器,分类的准确率只有75.6%,融合光谱曲线特征后用SVM进行分类,准确率可达99.2%。此外,本文还计算了被检索样本和标准等级样本特征向量的欧氏距离,完成了腊肉新鲜度等级的检索任务,显示了该方法的可行性和有效性。
Abstract:
Taking the bacon as the experimental object to establish a method of classification and retrieval of meat food with the characteristics of spectral curve and image,and using ten 3×3 convolution layers,three 5×5 convolution layers,five 5×5 pool layers and two full concatenation layers model to feature the hyperspectral image,and the cross entropy was used as the optimization target. The multiple scattering correction and principal component analysis were used to preprocess and feature extraction,then the two features were fused,and the fusion features were classified by Support Vector Machine(SVM). The experimental results showed that only the image features obtained by CNN were considered as input data,and classified by SVM,the accuracy was 75%. However,when the features of hyperspectral curves were fused into the image features and classified by SVM,the accuracy could reach 99.2%. In addition,the Euclidean distance algorithm was used to return the ranked scores for the retrieved samples,which realized the retrieval of freshness levels. The experimental results showed the feasibility and effectiveness of the proposed method. In addition,the Euclidean distance of the retrieved sample and the standard sample feature vector is calculated,which realized the retrieval of freshness levels,which showed the feasibility and effectiveness of the method.
百度学术
下载:
