基于卷积神经网络和高光谱成像技术的芒果成熟度判别

Determination of fruit ripening degree based on convolutional neural network and hyperspectral imaging technology

摘要:

为了无损判别芒果成熟度，本文着力探索利用可见近红外高光谱成像技术与深度学习方法结合。通过漫反射高光谱成像系统，采集 400~1000nm 波长范围内未熟、成熟和过熟芒果的高光谱数据，运用SPXY算法按 6:2:2 划分为训练、验证和测试集。经小波阈值去噪（Wavelet Threshold Denoising, WTD）、多元散射校正（Multiplicative Scatter Correction, MSC）等光谱预处理后，构建一维卷积神经网络（1-Dimensional Convolutional Neural Network, 1DCNN）模型。同时，采用主成分分析（Principal Component Analysis, PCA）和自举软收缩（Bootstrapping Soft Shrinkage, BOSS）提取特征波段，搭建偏最小二乘判别分析（Partial Least Squares-Discriminant Analysis, PLS-DA）与极限学习机（Extreme Learning Machine, ELM）的分类模型。对比分析表明，基于多元散射校正预处理的 1DCNN 模型性能最佳，在测试集上分类准确率达98.30%，Kappa系数为0.97。研究结论表明，高光谱成像技术与神经网络算法相结合，能高精度无损判别芒果成熟度，为芒果自动化分拣的实际应用提供理论基础。

Abstract:

In order to non-destructively discriminate mango ripeness, this paper focuses on exploring the use of visible near-infrared hyperspectral imaging combined with deep learning methods. The hyperspectral data of unripe, ripe and overripe mangoes in the wavelength range of 400~1000nm were collected by diffuse reflectance hyperspectral imaging system, and the SPXY algorithm was used to divide them into training, validation and testing sets according to 6:2:2. After the spectral pre-processing such as Wavelet Threshold Denoising (WTD) and Multiplicative Scatter Correction (MSC), a 1-Dimensional Convolutional Neural Network (1DCNN) was constructed. At the same time, Principal Component Analysis (PCA) and Bootstrapping Soft Shrinkage (BOSS) were used to extract the feature bands, and the classification model of Partial Least Squares-Discriminant Analysis (PLS-DS) and Extreme learning machine (ELM) was constructed. Comparative analysis shows that the 1DCNN model based on multiple scattering correction preprocessing has the best performance, with a classification accuracy of 98.30% and a Kappa coefficient of 0.97 on the test set. The conclusion of the study shows that the combination of hyperspectral imaging technology and neural network algorithm can discriminate mango ripeness with high accuracy and without loss, and provide a theoretical basis for the practical application of automated mango sorting.