针对传统垩白大米检测主观随意性大、 可重复性低、 检测过程耗时费力、 准确率低等问题, 提出一种基于可见光谱图结合深度学习算法的垩白大米检测手段。 用CCD彩色摄像机获取垩白大米和正常大米可见光谱图, 对图像进行旋转、 翻转以及调整对比度等随机图像变换方式提升网络训练数据集, 防止深度检测模型在学习过程中出现过拟合现象。 构建了7层深层次卷积神经网络模型, 包括卷积层、 池化层、 全连接层和输入输出层, 通过网络模型对采样的大米可见光谱图集进行卷积与池化操作, 采用迭代学习训练方法获取大米可见光谱图在卷积层输出的特征参数, 采用连接非线性ReLU激活函数来降低训练时间, 以加速大米可见光谱图有效抽象特征提取的收敛速度； 然后将深度神经网络嵌入池化层, 对大米特征降维以获取能够表达正常大米和垩白大米可鉴别显著意义特征； 最后在全连接层输出进行分类, 从而实现对垩白大米的精确识别。 基于可见光谱图的大米垩白深度检测方法比传统基于可见光谱图的垩白大米鉴别特征提取方法免去了复杂的特征提取步骤, 由于卷积网络提取的特征对特定目标具有更鲁棒的表达, 算法精度较高且复杂度比较小, 泛化效果更好, 获得识别精度达到90%, 比基于传统特征提取的垩白大米鉴别方法识别精度高, SIFT+SVM, PHOG+SVM和GIST+SVM模型识别精度分别为70.83%, 77.08%和79.16%。 提出的方法为当前我国现代农业生产中实现大米品质自动化快速精准检测提供了理论依据和有效的技术手段, 对于现阶段实现大米品质人工智能检测产生实际意义。

Aiming at the problems of subjective randomness, low repeatability, being time-consuming and low accuracy of traditional chalky rice detection, a new method based on visible spectrogram combined with deep learning algorithm is proposed to meet the requirement of rapid and accurate rice quality parameters in modern agricultural production. In the experiment, CCD color camera was used to obtain the visible spectra of chalky rice and normal rice. Random image transformation methods such as rotation, flipping and contrast adjustment were used to enhance the network training data set to prevent the fitting phenomenon of the depth detection model in the learning process. In this paper, seven deep-level convolution neural network models, including convolution layer, pooling layer, full-connection layer and input-output layer is constructed. The visible spectral image of rice is convoluted and pooled by network model. The characteristic parameters of visible spectral image of rice in convolution layer are obtained by iterative learning training method. The non-linear ReLU activation function is used to accelerate the convergence rate of the effective abstract feature extraction of rice; then the pool layer is employed to obtain the distinguishable semantic features that can express normal rice and chalky rice; finally, the data are transported into the full connection layer. The chalky rice can be identified accurately by classification. The method of rice chalkness detection based on convolution neural network eliminates the complicated steps of feature extraction compared with the traditional method. Because the features extracted by convolution network have more robust expression for specific targets, the algorithm has higher accuracy and less complexity, and the generalization effect is better than the traditional method based on visible spectrogram. The recognition accuracy is up to 90%. The recognition accuracy of SIFT+SVM, PHOG+SVM and GIST+SVM are 70.83%, 77.08% and 79.16% respectively. The proposed method provides a theoretical basis and effective technical means for the realization of automatic and accurate detection of rice quality in modern agricultural production. Therefore, this study has certain theoretical value and practical significance for the realization of artificial intelligence detection of rice quality.