提出一种基于可见光谱图多模态词典特征低秩稀疏表示框架的大豆外观品质判别方法,以精确确定大豆品质等级。首先,提取大豆粒子可见光谱图像的多尺度空间梯度特征和色差分量(YCbCr)颜色空间特征;将上述提取的空间梯度特征和颜色空间特征看作视觉词汇,通过Kernel K-means聚类算法获取视觉词汇的核空间局部分布聚类中心,形成视觉词典;然后,使用低秩稀疏表示法耦合上述两种特征,用于消除高维异质模态词典描述符中冗余信息的影响;最后,在高维耦合空间中根据样本之间的度量对低秩稀疏耦合表示多模态词典特征进行分类。所提方法充分利用多模态多尺度空间梯度特征和YCbCr颜色空间特征来描述大豆粒子外观品质的语义特征归属。实验结果表明:建模集和预测集总的识别精度分别达92.7%和80.1%,所提方法的识别精度优于文献中提出的基于单一模态的视觉词典特征表示方法。

A method for discriminating the appearance quality of soybeans based on the low-rank sparse (LRS) representation frame of multimodal lexicon features in the visible spectrogram is presented to accurately determine the soybean quality level. Firstly, multi-scale spatial gradient features and YCbCr color space features of the visible spectrogram of soybeans are extracted and regarded as visual vocabularies. The Kernel K-means clustering algorithm is used to form the local distribution cluster center of visual vocabularies in kernel space,thereby generating a vision lexicon. Secondly,the LRS representation method is used to couple the two type of features, thereby eliminating the effect of redundant information in high-dimensional heterogeneous modal dictionary descriptors. Finally, the LRS representation coupling multi-modal dictionary features are classified according to the metric between samples in the high-dimensional coupling space. The proposed method makes full use of multi-modal and multi-scale spatial gradient features and YCbCr color space features to describe the semantic feature attribution of appearance quality of soybeans. The experimental results show that the recognition accuracies of training set and prediction set are 92.7% and 80.1% respectively, and the discrimination accuracy of the proposed method is better than that of single-visual-mode based vision lexicon feature representation method.