基于贝叶斯优化XGBoost的小麦粉麸星含量预测

    Prediction of bran speck content in wheat flour based on XGBoost with Bayesian optimization

    • 摘要: 针对基于传统图像处理的小麦粉麸星含量测定存在分割困难和统计误差大的问题,提出基于贝叶斯优化极限梯度提升算法 (eXtreme Gradient Boosting, XGBoost) 的小麦粉麸星含量预测模型。使用方差膨胀因子 (Variance Inflation Factor, VIF) 对小麦粉RGB图像的颜色和纹理特征进行筛选,构建小麦粉麸星含量数据集。将筛选后的特征和麸星含量参数分别作为模型的输入和数据标签,训练完成后,将其与随机森林(RF)、梯度提升决策树(GBDT)、AdaBoost (Adaptive Boosting)和卷积神经网络(CNN)进行对比试验。结果表明:XGBoost的平均绝对误差为0.004 4,与RF、GBDT、AdaBoost和CNN相比,分别降低了18.51%、22.81%、18.51%、24.14%。该算法模型可以实现麸星含量的准确预测,对指导小麦粉适度加工、提高检测效率具有较强的实用价值。

       

      Abstract: Aiming at the problems of segmentation difficulty and significant statistical errors in determining wheat flour bran speck content using traditional image processing methods, a predictive model for wheat flour bran speck content based on Bayesian-optimized eXtreme Gradient Boosting (XGBoost) algorithm was proposed. The Variance Inflation Factor (VIF) was utilized to perform feature selection on the color and texture characteristics of wheat flour RGB image, thereby constructing a dataset for wheat flour bran speck content. The selected features and bran speck content parameters were used as input and data labels of the model respectively. After training, comparative experiments are conducted, with random Forest (RF), Gradient Boosting Decision Tree (GBDT), AdaBoost (Adaptive Boosting), and Convolutional Neural Network (CNN). Experimental results indicated that the mean absolute error of XGBoost is 0.004 4, which represented a reduction in mean absolute error by 18.51%, 22.81%, 18.51% and 24.14% compared with RF, GBDT, AdaBoost, and CNN, respectively. This algorithm model can accurately predict the bran speck content, offering significant practical value for guiding the moderate processing of wheat flour and enhancing detection efficiency.

       

    /

    返回文章
    返回