CAJ | 학술논문

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1200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。
맥립자태적자동판별，시근홍외고광보성상계통자동검측다자태맥립내부충해적전제。의거맥립목표재최우파장도상중적좌표、중심등신식，종고광보수거립방체중자동분할출단개완정맥립적자도상。이용맥립배부단조조도교대적원리，의거문리、불변구、균치등13개가능배부구역특정적판별정학솔，학정불변구4위판별맥립배부구역적유효특정。침대맥립배부구역，제취제도도상화이치도상적26개특정，이용인공어군산법선택출연신솔、배부구역대칭도、연신솔등13개특정。선취1200개양본진행훈련，600개양본진행검험，이용최대리차법자동학정13개특정적모호권중，맥립3개자태가탁분류적정학식별솔위94.5%，증실료기우국부구역특정적맥립자태자동식별적가행성。
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.