基于胚部区域特征的麦粒姿态自动识别

Automatic posture recognition of wheat kernels based on germ features

麦粒姿态的自动判别,是近红外高光谱成像系统自动检测多姿态麦粒内部虫害的前提。依据麦粒目标在最优波长图像中的坐标、重心等信息,从高光谱数据立方体中自动分割出单个完整麦粒的子图像。利用麦粒胚部端粗糙度较大的原理,依据纹理、不变矩、均值等13个可能胚部区域特征的判别正确率,确定不变矩4为判别麦粒胚部区域的有效特征。针对麦粒胚部区域,提取梯度图像和二值图像的26个特征,利用人工鱼群算法选择出延伸率、胚部区域对称度、延伸率等13个特征。选取1 200个样本进行训练,600个样本进行检验,利用最大离差法自动确定13个特征的模糊权重,麦粒3个姿态可拓分类的正确识别率为94.5%,证实了基于局部区域特征的麦粒姿态自动识别的可行性。

The insects in internal kernels not only harm the grain directly but also downgrade the market value of grains. It is very important to detect insects inside grain kernels accurately. The image features extracted from the kernel with different postures was a large difference. The automatic posture recognition of wheat kernels was primary to automatically detect insects inside kernels based on the near-infrared hyperspectral imaging technology. The five kernels were placed in a black plastic plate, and were not touched at the same posture. The hyperspectral data cubes of the wheat kernel with three postures were separately acquired at the first 14, 18, and 23 days after the rice weevil oviposition. The three postures of the kernel were crease down, crease up, and crease side. The hyperspectral data cube of the kernel was analyzed by the principal component analysis, and the optimal wavelength was 927.61 nm. According to the coordinate, center of gravity, and other information of the kernel in the optimal wavelength images, the sub-image of a single wheat kernel was automatically segmented from the hyperspectral data cube. There were two possible germ regions at both ends of the kernel, and the roughness degree of the germ region was larger. Thirteen features of the possible germ region, such as texture and invariant moments, were extracted to characterize the roughness degree. Each of the 13 features was used to discriminate the germ region by the principle of the germ region having a larger roughness degree. The recognition accuracy rate was 100%using the fourth invariant moment to identify the germ region of kernels, so the fourth invariant moment was the optimal feature because of the highest recognition accuracy rate. The six texture features and the seven invariant moment features were extracted from the gradient image of the kernel germ region. The thirteen features such as extension ration and symmetry degree of wheat germ region were extracted from the binary image of kernel germ region. The twenty six features were extracted in order to character the differences of the three postures of kernels. The evaluation principle of the feature subset was proposed based on the recognition accuracy of the v-fold cross-validation training model and the number of the selected features. The artificial fish swarm algorithm was applied to the feature selection of the kernel postures. The algorithm selected 13 features that composed the optimal feature space from the 26 features, such as symmetry degree of wheat germ region and complexity. The recognition accuracy of the validation set was up to 93.33%. According to the three-fold standard deviation principle of normal distribution, the classical matter-element matrix and the extensional matter-element matrix were constructed with the feature mean values and the standard deviations. The quantitative fuzzy feature weights were automatically determined based on the maximum deviation fuzzy analysis. The 1 200 samples were selected to train, and the 600 samples were used to validate. The three postures of wheat kernels were recognized by the extension classifier based on the fuzzy weights. The 567 samples in validation set were correctly identified, and the recognition accuracy of the classifier was 94.5%. The experiment showed that the automatic posture recognition of wheat kernels based on the local region features was feasible.