基于高光谱成像和多重优化策略的橙子早期腐烂检测

Early Detection of Orange Decay Based on Hyperspectral Imaging and Multiple Optimization Strategies

摘要:

针对橙子早期腐烂识别中传统方法检测精度低的问题,提出了一种融合高光谱成像与多重优化策略的检测方法,旨在实现早期腐烂橙子的有效检测。基于450 ~ 1 050 nm波段的高光谱数据,通过Borderline-SMOTE算法合成增强数据(KL 散度为0. 02,Wasserstein 距离为3. 4),采用ReliefF算法筛选了24个关键特征波长,结合贝叶斯优化构建机器学习及CNN分类模型,并系统地评估了模型分类性能与计算效率。CNN模型经贝叶斯优化和特征筛选后分类误差降至0. 008,运行时间由910. 4 s 减少至177. 9 s(降低80. 5% ),测试集准确率达99. 0% 。

Abstract:

A detection method that combined hyperspectral imaging and multiple optimization strategies was proposed to effectively detect early decay of oranges, addressing the issue of low detection accuracy in traditional methods. Traditional detection approaches, such as visual inspection and manual sorting, were highly subjective and fail to identify subtle changes in early decayed oranges, leading to substantial post-harvest losses in the citrus industry. To solve this problem, the hyperspectral data in the 450 ~ 1 050 nm wavelength range were collected by using a professional hyperspectral imaging system, covering the visible and near-infrared regions closely related to fruit internal quality. Then, enhanced data were synthesized by the Borderline-SMOTE algorithm with Kullback-Leibler (KL) divergence of 0. 02 and Wasserstein distance of 3. 4, which effectively alleviated the sample imbalance problem between healthy and early decayed oranges. Subsequently, totally 24 key characteristic wavelengths were screened out by the ReliefF algorithm to eliminate redundant information and reduce computational complexity. Machine learning based classification models and CNN model were constructed in combination with Bayesian optimization, which optimized key hyper parameters to improve model performance. A systematic evaluation was carried out on their classification performance and computational efficiency. After Bayesian optimization and feature selection, the classification error of the CNN model was reduced to 0. 008. The running time was significantly decreased from 910. 4 s to 177. 9 s, representing a reduction of 80. 5% . The accuracy rate on the test set reached 99. 0% . The research result can not only provide a reliable technical solution for early decay detection of oranges, but also lay a theoretical foundation for the development of rapid and intelligent detection equipment, which was of great significance for promoting the high-quality development of the citrus industry.