基于多模型特征融合技术的皮蛋凝胶品质分级方法

Quality Grading Method of Preserved Egg Gel Based on Multi-model Feature Fusion Technology

摘要:

为了解决工厂在皮蛋凝胶品质分级时存在的成本高、主观性强等问题，设计了一种适用于可见/近红外光谱数据的多模型特征融合分类模型。使用CARS（Competitive adaptive reweighted sampling）算法提取皮蛋可见/近红外光谱的特征波段；设计了一维高效通道注意力(One-dimensional efficient channel attention, ECA_1D)模块，并添加在所建立1DCNN模型中得到了1DCNN_ECA模型，提取光谱的卷积特征；在LSTM网络中引入自注意力(Self attention)模块建立了LSTM_Self模型，来捕捉光谱序列的长期依赖关系；融合CARS算法、1DCNN_ECA模型和LSTM_Self模型所提取的特征建立了特征融合模型TripleFusion模型，其分级准确率可达95.0%。研究结果表明多模型特征融合方式可以弥补单模型在特征提取能力上的不足，大幅提升模型的分类准确率，解决了皮蛋凝胶品质分选难题，丰富了可见/近红外光谱数据的分析和建模方法。

Abstract:

Aiming to address the issues of high cost and subjectivity in the industrial grading of preserved egg gel quality, a multi-model feature fusion classification framework tailored for visible/near-infrared spectral data was proposed. Firstly, key spectral wavelengths were extracted by using the competitive adaptive reweighted sampling (CARS) algorithm, and a support vector machine (SVM) model was built, achieving a classification accuracy of 90.8%. Secondly, a one-dimensional efficient channel attention module (ECA_1D) was designed and integrated into a residual-connected one-dimensional convolutional neural network (1DCNN), resulting in the 1DCNN_ECA model, which achieved an accuracy of 92.8% by extracting deep spectral features. Additionally, a long short-term memory (LSTM) network was enhanced with a self-attention mechanism to construct the LSTM_Self model, effectively capturing long-range dependencies in spectral data and reaching an accuracy of 92.1%. These three feature representations, derived from the CARS algorithm, the 1DCNN_ECA model, and the LSTM_Self model, were further fused to develop the TripleFusion model, which achieved a grading accuracy of 95.0%, outperforming all dual-model fusion configurations. The results demonstrated that multi-model feature fusion can compensate for the limitations of individual models in feature representation, significantly improving classification performance. This work can effectively address the challenge of non-destructive grading of preserved egg gel quality and provide a novel and robust approach for visible/near-infrared spectral data analysis and modeling.